instruction
stringclasses 1
value | input
stringclasses 1
value | output
stringlengths 0
113k
|
---|---|---|
write a youtube script | I like planes I like boats I like planes I like boats. but what if I want that boaty planess in one?! oh no! I like planes and I like boats these are the words that Xyla Foxlin will say every morning! so anyways Its YouTube makers secret santa time yes! Loads of makers all group together we all make stuff for each other weve been doing it for like three years in a row now. This year I have Xyla Foxlin now she lives in LA thats gonna cause me problems later and basically you know from her videos shes made a boat before and I know shes got a pilots license from her social media stuff So I have a plan okay unfortunately were not making a flying boat because it is the th of December these videos are supposed to go on the th of December I have to post whatever I make to Los Angeles. I havent made it yet oh yes Im behind. Right so what Im thinking boats have oars we know boats have oars planes have propellers but boats can sort of have propellers as well in the term of props you know outboard motors. So so Im thinking well put propellers or outboard motor things in the oars. Genius Furze! Now Im not sure how its going to work because obviously when youre in a boat you push the Oar into the water and then you pull against the water and push the boat forward but if the oar is trying to push itself through the water does the oar just try and get out of your hands and disappear? I dont know but I cant think of anything better so lets make it! Ok Ive hit simply bearings hard its where I get all my bearings chains and gears from. This is not sponsored theyre just a nice company. So anyway I think Im going to go with this chain which is like mm chain so what Im thinking Colin what have you bought chain? Well basically oar propellers set in there chain around the shaft with a cog on it down through the um the pole the oar thing and then to some sort of central Hub where there will be a motor some some gears and stuff that will spin them both so they will both go around off the one motor. Right so Im going to get all this set out make myself some little templates do a little design just get on with it Colin! Its half ten now you got to get on with it! what is this.... Xylas arm slashing hair tangling super paddle... oh middle bit! Right thats the center bit now this middle Cog which is just over a bolt at the moment thatll be on a shaft which will be attached to whatever power source we choose. Now this chain goes up and over it back and forth and then you can see if I pull it look if the chains nice and tight it should go across the top of it and then pull back underneath it in equal amounts and therefore spinning the propellers at each end. Now the chain itself will slide down tube this is going to be noisy Colin yes it is! and then theyll go into these slots like that. both ends thats that right theres another plate going to go on top of that and then Ive got to weld them onto the things and hope that I can get to everything. So many complications. Right thats welded on. Cant get to them cogs very easily now. check. right the ends propellers now Ive started us off look Ive put some little thicker plates on the end. Ive welded them on and then whatever we make we can then bolt onto that and if its wrong or breaks we can then unbolt it. So we dont have to grind it off or cut it off or anything like that right so the ends about that! This is my concept people... yes we still need something that resembles a paddle even though weve got our propeller so Ive made this obviously its got a lot of holes in it or else youre just propellering against your paddle. Bolt this on here chain comes out round axle axillo skinny. but because its a little bit floppy Im gonna put like a secondary one on which will slide on bolt through it stick it off make it double skinned like the middle bit itll be a bit stiffer... yes! OK thats got it all together that was a nightmare. Now then its all connected up it spins around both go around in the same direction of course which is good now in the middle I have to shave a bit off the gears which is a complete pain cuz I had to try and get them out and end up in the end like plasma cutting some of my own spanners out of some . mil Steels we had to get down and slotting in but anyway its all good. Now the chain is tensioned via these little slots here so basically you put it all together and you just pull the whole thing back and tighten it up. That is good. Now the next thing we need to consider Colin is the power source so anyway I thought what did I think? Right and so I was thinking we need a nice motor thats got power and got speed I always found its quite a tricky thing to get hold of and I thought well wait a minute why dont we do something that she might already have and I was thinking... drill shell have a drill wont she! I dont think a drills fast enough but I know what Ive got thats a bit quicker grinder! oh!! haha! Oh God! its a bit quick! I think well use a grinder! Weve got no choice! Right lets somehow rig up a grinder on this thing and then see how much of a laceration machine I have produced! I cant help noticing this is all starting to look a bit big and a bit heavy! I dont think shes that big you know? No I think itll be all right shes not that small Ive seen her stood next to a rocket shes pretty tall actually! whatever whatever! Right... see what happens here! oh hahaha! If that sound does not translate onto video that is like belt of knives X I dont even think thats full belt! I think I can manage this. Right this sounds absolutely horrific and everything but does it actually work? Ok here we are in December down the river its freezing! Right Ive rigged up like this really crude trigger system put a drill on it just for the time being .Im gonna give that a test first. Right gonna put my coat back on well have a go. The only the only thing Ive got to attach the actual GoPro to is the oar. So the footage for this might be a little bit wobbly. Immediately Im returning to this one thats not a good! Ive dropped it! why did I build something that goes into blooming water in December! faster and faster! oh no! This is a design problem the chain is bringing water up and dribbling it in my oh no Ive got a wet... There is so much wrong with this but it sort of works in a weird way. I mean it spins around in circles Ive now got wet A! Just thinking about this Id be better with the thing the other way around. We need to do this and you have the triggers all wrong! oh God this is terrible! Im bloody soaked! right right! You know what I think this actually works! This is working! Right Im going to stick the angle grinder on it because this things A the batteries are dying and its not as fast as it could be. oh this is... I cant feel my hands! Just get back in the old boat! Right people oh my God check this out! ah! oh my God Im so bloody wet! The grinder is not quite powerful enough in terms of it spins too fast and hasnt got the torque but the paddle does work. okay we do have some success here. Thats it for today its cold Ive had enough! Now despite getting freezing cold and soaking wet there is some potential in this! Basically the grinder spins very fast its not quite got the torque it needs to as soon as you put it in the water it basically stops and the drill wasnt actually as bad as I thought it was going to. So Im thinking Im gonna go back for another attempt with some proper clothing on and a better drill with a bigger battery on ! Right back again look like Im searching for dead bodies but at least Ill be dry. Come on this is gonna work! Here we go people! this is working! If you put it in look it will work like an outboard motor and spin me around! oh my God its actually icy! t more speed for more effort! back to the workshop! We may not have invented a new extreme sport but there is potential in this. If you got the right motor built into the center fixed it in properly so it dont spin around and smack in the head. Put the batteries either side of it so you can hold the thing properly and it wasnt quite so heavy I think thered be a bit of potential here Now we aint got time to do that because this videos got to go on the internet tomorrow and obviously clearly Im not going to get all that done. Now Colin if its going on the internet tomorrow arent you supposed to be sending this to Zyla? yes I am but I did decide well a few days back now that this is probably not the best thing to just hand over to somebody for them to try with all the spinning blades of dome potential lacerations. Ive made her something else and sent that in tribute to this this is a dedication to her and her love of planes and boats youll have to go and see what the other gift is I thought it was a lovely and wonderful gift but now Ive thought about it and Ive looked at it since... it does look a bit kinky... so youll have to go and make your own minds up! hmm! But this is YouTube makers secret santa its not just about what Im making for other people people have made for me as well and I have mine here! lets undo it. OK its from Emily the Engineer now I know this because it says here on the box who its from. Now this will be interesting. She is like a Marvel addict she does a lot of stuff with Iron Man suits shes made her own Iron Man suits and stuff so Ive been quite interested to see what this is. Rightyho! Got notes Colin Merry Christmas I know you often enjoy wearing neck ties so I thought Id make you a device to make putting them on a little more interesting! So it is a tie launcher for launching ties onto me! this is what you gotta like about YouTube makers Secret Santa you just never know what youre gonna get but its very well wrapped. oh you star Emily look shes even got the old Safety Tie warning triangle on there. Right shes advised I try it with a mannequin first so Ive got Mrs yellow in its got a little laser pointer on the top so its put like a little dot where you want to fire the tie I like it! Right lets put a little COT cartridge in the back which is what its powered by and give it a test. Here we go. oh Ill stand in front of that! Right no more Mrs thingy come on Furze tie to the face! By the way Emily this gets a thumbs up from me! Right so take my tie off collars up collars up! Ready to accept! Come on timing. I have no idea if that nearly got me at all oh thats so close Yeah! it works!!! Fantastic! ah! That is brilliant! Right now of course YouTube Makers Secret Santa the whole point the whole reason we do this is so you can go on your little course going from one video to another so you can even go backwards you can go and see how Emily made this or you can go forwards and go and see what Zyla has made and go and see her reaction to her nonpresent because Ive kind of Ive kind of failed a little bit this year but anyway we dont think it was it but anyway . so go on the trail I think theres of us in it this year you know the reason why I do this is to try and spread the love to all the other makers so just go and check them all out things like this absolutely fantastic thanks for watching hope youve had a good year see you next year where of course there is more tunneling to do weve got a car park were going to connect it to the bunker and theres all sorts of things to do at the farm as well. boom for you |
|
write a youtube script | Im making a functional tabletop bowling out this is going to be fully functional with the working pin Setter to start off I started by printing some pins relative to the scale of my in marble bowling ball the first subsystem I want to tackle is the lane and bumpers I think itd be cool to have a Deployable bumper system for those who arent so skilled I knew I needed some sort of locking mechanism that could be pulled up and stopped at the top of its position this is a locking parallelogram mechanism as the mechanism is lifted a pin rides along a chamfered edge to lock it into its top position the mechanism cant be moved until someone pulls a springloaded latch to release the mechanism and fold it back down I recreated my own version of the mechanism in solid works to fit the scale of my bowling lane and printed out the parts for testing my first design didnt work well but it never does its just for learning purposes I kept reiterating and adding small improvements with each print for instance the springloaded latch was a bit too Wiggly and needed more walls to contrain its movement it took me about six tries to get this perfect where the mechanism can be pushed into place and now it has a perfect satisfying release to fold it back down follow to see the rest of my build |
|
write a youtube script | this is jaws jaws is not a nutcracker hes a nut obliterator hes times more powerful than a normal nutcracker so load up and strap on your body armor because with jaws you just might need it Music jaws was made because i thought it would be really funny to crack this kind of nut so well be testing that and of course sticking everything else we can think of in there if youre wondering why something like this should exist i mean come on its gonna be awesome three two one wow this all started when i made an explosive powered bat which as best i can tell broke the mlb home run distance record ever since ive been wanting to make a nutcracker using this technology and tis the season the bat was powered by these little explosive blanks which are used to drive nails directly into concrete they should be really good at driving nutcracker jaws into everything the red tip on these blanks means that theyre maximum power which is really cool but you know whats even cooler four blanks at the same time getting this thing working was pretty hard it really wants to just violently tear itself apart that reminds me i need to tell you do not try this at home if you dont know exactly what youre doing this is incredibly dangerous and even deadly theres a lot of engineering testing and precautions that im not even showing in the video that im doing to make this safe so live vicariously through me and if the friend asks you if you want to make something powder actuated just say no all right lets go back to the beginning a traditional nutcracker works by amplifying the force from your hand with a lever to squeeze a nut really hard and crack it with powder actuation i should be able to make a nutcracker that generates up to pounds of force thats like times stronger than this nutcracker and thats all thanks to these wonderful little blinks youre supposed to put them in a special tool which tickles them in just the right way they shoot out super high pressure gas which drives a nail into concrete my goal is to somehow use this expanding gas to actuate the nutcracker although im not going to use a lever because they really like to snap and besides this is so powerful i dont need no stinking lever if i put a piston over these blanks the expanding gas will shoot it up and crush these nuts this is similar to how my bat worked but theres a big difference this is the pushy bit from the bat that launches the ball its cal because the pistons are half of an inch diameter and this is what im going to be making the piston for the nutcracker out of cal inches this is going to make the nutcracker generate about times the force of the bat to understand why we have to talk about pressure pressurized gas is basically gas that wants to take up a lot of space smooshed down into a tiny little space and like a compressed spring it wants to expand except the gas wants to expand in every direction at the same time pressure is measuring how hard the gas is pushing on a given area if i have a bigger area it will push harder the problem is that when you start making the piston really big it really wants to tear itself apart and thats because the stress in the cylinder grows like this as you make the piston bigger so to try to make it strong enough i could just make everything thicker than a bowl of oatmeal but that isnt a very good strategy for one i could be wrong and it could explode in my face but also metal is really expensive and if i make it thicker than it needs to be im wasting material which is a tragedy this is why engineering is a thing if i can calculate how strong the nutcracker needs to be and how strong a given design is i can design it to be just strong enough and its just like they say any old schmuck can build a bridge but only an engineer can make a bridge that barely stands im planning to bolt together a stack of thick steel plates to make the piston when the pressures pushing the piston up its also pushing downward the opposing forces will try to rip the nutcracker in half and the bolts are the only thing keeping it together so its critical that the bolts are strong enough so lets talk about steel if i tighten the nut on this big bolt its hard to see but the bolt is getting longer and if i loosen the nut it springs back to its original length usually at least if youre anything like me youve cranked the nut to the point where it suddenly gets easy to turn and whats happening is that the bolt stretches so much and it starts to permanently deform we need to use bolts that are strong enough but the question is what size different size bolts take different amounts of force to stretch but if i take the force it took to stretch any given bolt and divide it by the bolts cross sectional area i always get the same number which is interesting and this number is actually a really important property of metals its called yield strength its telling me how many pounds per square inch of crosssection steel can take before it stretches i can use this to directly calculate what size bolt i need and this is one of the ways that engineers predict how somethings going to behave rather than having to physically test it so at this point we have a piston that were going to shoot upwards utilizing the farts of these shells the plan is for the piston to be the nutcrackers torso so were going to move the shells over to the nutcrackers back where theres also room for the firing pins and a hammer and all that stuff when it goes bang all the gas shoots down the tube and launches the piston upwards but once its done i need to let the gas out i could have a little hole that opens up like this and lets the gas out but this is going to be incredibly loud and nutcrackers are supposed to be used in the family room with everyone gathered around them loud explosions would totally spoil the mood and besides everyone knows that the only loud explosions allowed in these situations is vicious political debate so im going to add a tube to the piston that gets pulled up as the nutcracker fires when the jaw reaches the closed position itll open up a little hole that lets the gas shoot down the tube which will have a muffler on the end of it which will make it quiet this tube lets me solve another big problem which is stopping the piston remember its going to be a chunk of inch steel and if i fire the nutcracker with nothing in it its going to be going up to miles per hour if it slams into a solid steel wall at that speed it will destroy everything so im going to put a spring right here which has to be compressed before the piston hits the wall this will bring it to a gentle stop and by gentle i mean gs which would liquify your brain but its much better than hitting the wall this is the spring were going to be using the technical spec for this is beef supreme remember that gas that tried to rip our bolts in half well it also pushes out sideways which is a big problem we need to make the nutcracker thick enough so that this doesnt happen but how thick is thick enough just like the bolts our goal is to make sure that this tube never stretches to the point that it permanently deforms but figuring this one out is tricky if i pressurize the pipe and look at how much its stretching it varies in a complicated way across the pipe this problem is a lot easier to think about if we zoom way in on the pipe and look at a little tiny piece thats so small the variation across it is negligible we can imagine what happens to this little piece the pressure in the tube is pushing it outwards but this piece has little pieces next to it that pull on it to keep it from moving away it also has a piece behind it thats going to keep it from moving which is also held in place by pieces next to it and so on everything is pushing and pulling and squeezing on everything if this seems like a huge complicated mess youre right ill spare you the specific details but math makes it possible to precisely specify all these different things that happen in a little piece which you can use to figure out the stretch or the stress for any part in the tube then all we have to do is run the numbers until we find a tube thats thick enough so that it doesnt explode if youve ever wondered why anyone cares about calculus or differential equations this is why they are incredibly powerful tools for a huge number of problems i just ran the numbers and theyre kind of crazy to keep the nutcracker from exploding the walls have to be an inch thick which is really thick hopefully this reinforces why you dont want to mess with these things i would have never thought that the walls need to be that thick all we got to do now is design it ive been working on this for a couple days now and this is what ive come up with basically a giant block of steel which is what it takes to hold this thing together steel is pretty slow to machine so this is going to take a while if youve ever wondered what its like to make something like this look no further Music so Music man that took forever assembly is usually a lot easier than the fabrication this is the triggering mechanism these are the pins that ignite the shells Music all of these big steel parts are the piston i tried to reduce the number of seals but theres still a lot of seals all the nonstrength critical parts are d printed on the fuse one this is starting to look like a nutcracker i think hes ready to make his big debut Music sporting a classic military uniform with an avantgarde flare he has a face that says see what happens will the influencers of the world adopt his style i think that would be great but only time will Music tell all right lets be honest this is amazing this is so cool ah its huge theres a normal nutcracker i love this this is so good this little backpack that hes wearing is what holds the shells and does the firing and just like a real nutcracker you push this lever down to fire it to load it you open the breech and put the shells in here then when youre done you close the breech and then it and its armed my goal was to make it impossible to set this thing off when you dont intend to and theres a series of safeties to ensure this the first thing we have is the safety stick this physically prevents the nutcracker lever from being depressed it also goes inside of the nutcracker and physically blocks the hammer from reaching the shells so even if you somehow release the hammer it just cant go off if the slide isnt fully shut it blocks the hammer as well and the slide cant be fully shut unless the breech is fully locked this prevents it from being triggered with the breech partially locked or open which could be really dangerous and then if we decapitate him these bolts that hold the body together are designed to fail before the body itself this means that if somehow something went pop its gonna shoot this plate straight up rather than pieces of the body outwards which is much more dangerous thats a lot of engineering and precautions but i still dont trust my engineering that much so weve got body armor hard hat eye protection im cowering behind the barricade and it has remote start so i dont have to be anywhere near it a while back i had to fast for a couple of days and this is a reenactment of what it looked like when i finally got some food three two one this isnt nearly as powderized as i would have expected lets do something a little more quantitative metal nut versus two shells three two one what what some really precise measurements i think i can barely detect some squish but i shouldnt need calipers to see this it should be smooshed a lot more i spent a lot of time wandering in the wilderness and were just going to skip over that part this has just been driving me crazy ive been messing with this for multiple days now ive fired it a million times and it kind of works but the power just is not matching what it should be i thought there was a bunch of different problems i remade the chamber that holds the shells three times and im pretty sure i figured out the actual problem the gas from the shells comes out of this hole and goes into this hole to drive the piston and it would really hurt the performance of the nutcracker if these holes didnt line up no one could have guessed that right i mean come on it only took three tries to figure it out there is at least a quick hack to fix this three two one Music its better but it still isnt as powerful as i calculated the nut with three shells should be flat im just going to go for four i want to know if it can survive it i want to know if im gonna have to remake a bunch of parts three two one Music that was so loud even with the hearing protection i felt that in my soul it didnt even do anything with a nut its worse than three shells like what doesnt make any sense i dont understand what is wrong with you oh i see it looks like the pressures so high that it blew the tops of the shells off and squeezed them out into this little tiny gap in the breech which is surprising because the fit is really good on this its built like a swiss watch maybe a swiss watch made in china its pretty good unfortunately what happened here is pretty fundamental to the design of these parts i dont see any way to modify these to make it work i guess im going to go redesign and remake all these parts great its not that bad i just remade these parts and youre not going to believe it the firing pins dont line up with the shells so i get to remake them again all right it is that bad this doesnt fit this one should fit finally the new design bolts everything together so theres no gap to squeeze the shells into i just realized a big problem when this thing fires all that gas goes down through the muffler into the legs which are totally sealed it has nowhere to go i gotta find a way to get the hot gas out before i break something you know this is the best spot dont judge me Music all right i think hes finally ready for prime time and in this corner we have jaws weighing in at pounds he is a chonker and in the other corner is lil joe weighing less than two mice i have to say is this fair wait hes already devoured him is that even legal all right i think its working we get to do the fun part now we get to see what this little man is made of ive brought the wife in so she can render her opinions what do you think of jaws manicured mustache and eyebrows clearly had some dental work done theyre also metal hes got a grill and i really like his hair were going to start with what this thingy is designed for this is a rock hard macadamia nut what do you think is gonna happen to this nut its gonna explode and then the nut is gonna be smushed you heard it here first three two one wow Music it seems to have ejected all of the nut nut dust called it nice all right what do we got here some silly putty any predictions satisfying smoosh its not nearly as exciting it seems like it would be really fun boring snore i got jawbreakers but theyre too big red hot jawbreakers shatter it thats pretty much what i was expecting i think what about five jawbreakers same thing i ate too many jawbreakers oh i think im going to be sick oh disgusting its just all stuck to the roof of his mouth like peanut butter a glass marble this doesnt sound dangerous i just feel like everythings gonna shatter come on give him something hard a hardened steel ball bearing this is the hardest jawbreaker like split in half or something i dont know these bearings are really hard so were going with four shells im a little concerned about four if it does crack it could really come flying out fast so were gonna evacuate the wife get out of here doing four doing live three two one wow it just pressed the ball bearing into the jaws and the bottom the bearing is so hard its forming the steel jaws around it a literal jawbreaker it should be able to shatter this i think i just need harder jaws well save that one for later because right now were gonna do what weve all been waiting for metal nut versus four shells i think thats gonna get smooshed you think its gonna smush this three two one Music all right this looks great it looks like that cool s estee drawn middle school anyone whos cool in any grade it was one piece but with a little touch it snapped i cant believe it actually cracked the nut success my lifes work crack can i fire it uh yeah what do you want to smash what a battery i thought the same thing but thats a really bad idea do you have an extra lego men that i have i forgot that i had festivities very festive okay now we can lego man this little man went through santas court of law and has been deemed naughty oh yeah he goes into the nutcracker who is santas enforcement who is currently loaded with four shells two one Music Applause Music swap both arms all wow santa doesnt mess around dont get on the naughty list kids absolute carnage Music wow oh my gosh his poor head its like it split open in the back i think hes dead pretty crazy the big question i already married is do you think that this is an important invention a little i cant say no with him looking right at me all right hes not looking no i mean whats it gonna change the world id like to see it try see what she says behind his back its ridiculous i think hes very twofaced and very festive and very handsome do you like it yes you gotta ask the right questions this is my favorite kind of project its awesome arguably useless and definitely commercially not viable and that means that after all the engineering there isnt really anything for me to sell which makes projects like this really dependent on your support so im not trying to make you feel guilty if you enjoyed the video im happy but if you do want to help support projects like this there are a couple things that you can do so you might have noticed this awesome shirt that im wearing this is a blueprint of my unpickable lock and if you like it you can get one at stuffmatehere.shop another thing you can do is support these projects directly on patreon and the last thing you can do is take a minute to check out this video sponsor kiwico what kiwico does is every month they send you a crate that has everything that you need to do some kind of project and this one is an automaton it simulates santa coming to deliver all of your presents the reason i love these crates so much is they go beyond the project to teach you general concepts and in the case of this project santa is driven by a set of differently timed cams my unpickable lock actually had a bunch of cams people who see my projects like the unpickable lock ask me where did i learn to design stuff like that and it all started with a steady diet of kits and projects just like these theyre just a great investment in a kids future at this point pretty much all the kids in my extended family are getting kiwico crates if youre thinking about useful gifts that you could give a loved one you should check out kiwico and if you go to kiwico.com stuff made here theyll give you off your first month and they have a ton of different crates for different age groups and interests and thats it thank you kiwiko for sponsoring this video and thank you for taking the time to listen Applause Music Music you |
|
write a youtube script | Im making a D printed tabletop bowling alley and these are the lock in place bumpers the bumpers deploy and lock into its top position to release the lock you press a latch and it folds down I was pretty happy with this and I think the mechanism works well but at the end of the day when its locked its pretty Wiggly the mechanism hinges around these M bolts one way to reduce how Wiggly the mechanism is to reduce the size of the hole around the bolts look how many holes I have across this model and theyre all across different sketches so I made a variable called m three hole and that is linked to all the different holes so now I only have to change the diameter here and itll update all the holes across the entire model I reprinted the mechanism and now its way less Wiggly look at this I know that may seem simple or obvious but Im always looking for ways to make my modeling more efficient now you can lie and say it took three times longer than it actually did |
|
write a youtube script | The Gradient of Generative AI Release
Methods and Considerations
Irene Solaiman
Hugging Face
irenehuggingface.co
Abstract
As increasingly powerful generative AI systems are developed the release method
greatly varies. We propose a framework to assess six levels of access to generative
AI systems fully closed gradual or staged access hosted access cloudbased
or API access downloadable access and fully open. Each level from fully
closed to fully open can be viewed as an option along a gradient. We outline key
considerations across this gradient release methods come with tradeoffs especially
around the tension between concentrating power and mitigating risks. Diverse and
multidisciplinary perspectives are needed to examine and mitigate risk in generative
AI systems from conception to deployment. We show trends in generative system
release over time noting closedness among large companies for powerful systems
and openness among organizations founded on principles of openness. We also
enumerate safety controls and guardrails for generative systems and necessary
investments to improve future releases.
Introduction
What constitutes a robustly safe and responsible release of new AI systems from components such
as training datasets to model access itself urgently requires multidisciplinary guidance. There is no
overarching standard or standardsmaking body to form consensus for what constitutes responsible
release. This is particularly true for generative AI systems which can be leveraged for a broad
range of tasks and are sometimes referred to as generalpurpose . A subtype of foundation
models these systems generate outputs across modalities such as text and image. They can be
applied to both beneficial and harmful tasks. For example language models can be adapted to tasks
such as grammar correction and translation or be used for phishing and spam. The relative novelty
means new uses and misuses of these systems are still being discovered. Carefully considering
release strategies for present and less powerful systems better prepares and sets precedent for the AI
community and the broader affected community as increasingly powerful systems are deployed. Due
to the fast pace of AI progress and release developers deployers researchers and policymakers must
take action via community discussions guardrails and investments.
The current state of generative AI system release is largely determined by the organizations developing
systems. Developers are most likely to best understand the system but understanding impact and
ripple effects requires multidisciplinary expertise that is rarely housed in one organization. Waiting for
longterm evidence of consequences is infeasible for highrisk and powerful systems. A strictly closed
and vertical process to commercialization can lead to concentrated power among highresourced
organizations. An open process without ethical considerations can inflict and exacerbate risk and
harm from misuse to bias.
Many components make up a system throughout its lifecycle from training data to computing power.
This paper will primarily focus on the cumulative release of a model and its components by outlining
Preprint. Under review.
arXiv.v cs.CY Feb
key considerations in release release options along the gradient the timeline of released systems
and necessary investments to improve safe releases.
Previous Work
Discussion about safe release has been ongoing in the AI research community but there is no
standards body or default convener for these discussions. Initiatives such as the Partnership on AIs
Publication Norms for Responsible AI and Stanford Universitys Call for Community Norms
for the Release of Foundation Models have made space for discussion around the many options
for system releases and the many components involved and available platforms that have made
options complex.
At the core of release considerations is the tension between openness that shares rather than concentrates power and closedness that minimizes potential harm and risk. Broadly safety and risk
control development lags behind system development for example tools for detecting generated
outputs underperform as systems become increasingly powerful . For researchers in natural
language processing whether developers should be ethically responsible for downstream misuse of a
publicly released system is contentious with about half of researchers believing professionals should
be responsible.
Parallel fields such as opensource software deployment can share informative lessons such as the
ability for open source software communities to enable community research and crowdsourcing
work such as discovering vulnerabilities . Examining the specific use case of opensource
Deepfakes highlights the difficulty of managing downstream harms in realtime and the risk of safety
controls being seen as futile . While generative AI research would also greatly benefit from this
community insight lessons from software are not often directly applied due to substantial difference
in functionality . Ultimately personal values around openness are a large factor in decision
making and tensions can be further examined by output modality .
What is Being Released?
The parts of an AI system considered in a release can be broken into three broad and overlapping
categories
access to the model itself
components that enable further risk analysis
and components that enable model replication.
Components are organized based on their most straightforward use. There is overlap among these
components the same model cannot be replicated without its component for risk analysis such as its
entire original training data even if all replication components are available. Conversely components
for replication can also be analyzed for social impacts such as biases.
. The Model Itself
Access to the model itself includes the model weights and the ability to query adapt or otherwise
examine and conduct further research into a model. The range of access is expanded according to the
gradient in Figure .
. Components for Risk Analysis
These components are the parts of system development that could provide further insight into the
model the models capabilities the decision making process on what data was collected and how and
documentation of the process. Additionally this details system risks training data finetuning data
and information on people and human crowdworkers involved in adapting the model through methods
such as reinforcement learning with human feedback. This also includes evaluation results published
results from any evaluations that researcher and developers may have run on the base model. These
components may be withheld due to intellectual property IP rights consent or privacy concerns.
. Components for Replication
These components include a technical paper detailing the model training process and code used to
train the model as these can ease replication efforts. This also includes training information such as
configuration settings e.g. batch size and telemetry collected during training e.g. training loss.
These components may be withheld for competition IP and misuse reasons. They are also high risk
for misuse concerns as they can be repurposed or adapted to malicious or otherwise harmful use cases
.
Key Considerations in Release
Deployers should weigh the following considerations when making release decisions. Risks and
threats from increasingly powerful systems are difficult to enumerate and assess especially since
malicious actors and their incentives are constantly evolving . Taxonomies of ethics and risks of
specific systems can serve as a framework for potential harms. Specific considerations across
all generative systems are listed below.
. Concentration of Power
One of the most prominent arguments for providing access to systems is to avoid concentrating the
level of power that highresource organizations are collecting as one of the few groups capable of
developing and deploying these systems. Large technology companies are able to create powerful AI
systems because of their access to training data computing infrastructure and commercial capabilities
for deploying that system. This monopolization also gives these highresource institutions more
influence in AI development the behavior of these systems and the narrative and direction of the field
. Although these companies may provide access or even opensource their systems contributions
to system development are limited to people and resources working towards that companys interests
. Large companies are often geographically concentrated in Western countries whereas systems
are deployed globally which can asymmetrically impose cultural values . These companies can
also punish pushback or dissent . The people most affected and exploited by AI systems are
rarely found in large technology companies. They must be empowered to shape systems that also
benefit them or to opt out of interaction with AI entirely .
. Exacerbating Disparate Performance and Harmful Social Impacts
The fewer perspectives that are incorporated into the system development process leads to higher
likelihood the system performs disparately for different groups. AI systems can propagate harms
such as exacerbating social inequity and harmful biases which can be further
amplified in larger systems as scale increases . Means of measuring and mitigating risk in these
systems are largely cultural and contextdependent . The many technical and social aspects of AI
systems require robust research conducted with communities affected to ensure these
systems benefit and do not exploit marginalized groups if the systems are to be deployed among
these groups.
. Malicious Use and Unintentional Misuse
With more modalities of AI generation improving in output quality from high quality text to high
quality images the potential for harmful use cases also increases. Malicious uses such as the creation
of deepfake imagery AIgenerated disinformation and illegal and disturbing material
can cause severe emotional harm at the individual level and destructive institutional harm at the
societal level. Furthermore malicious actors have historically worked to circumvent safety
controls. Threat modeling will necessarily differ by modality but as systems improve in types of
outputs such as code generation potential harms can also broaden . While limiting access can
prevent some malicious uses and is often a suggested action to minimize misuse systems can
still be vulnerable to attacks with only querying functionality available .
. Auditability
The question of auditability addresses who is conducting audits and the level of access required to
effectively examine an AI system. Auditing must be considered both pre and postdeployment as
impacts from a system may not be detectable predeployment and when deployed impacts may be
difficult to trace back to a specific system . The actors conducting and capable of conducting
audits will likely require some level of technical skill even when numerous nocode tools are built. The
size of the system and its components also determine auditability the datasets that large generative
AI systems are trained on are not only difficult to analyze at scale but few tools exist to analyze large
static datasets . Formal audits alone cannot be the only insight or governance of a system .
. Accountability in Case of Harm
In the case that an AI system harms or is connected to harming people who or what is to be held
accountable is unclear. More open and deployed systems have a higher likelihood of a broader reach
and therefore a higher chance of harm. Since harm is not explicitly defined and not always physical
what constitutes harm can have a large range. The range may include encouraging physical harm
propagating social harms such as identity stereotypes and more abstract harms such as lack of access
to a system lowering opportunities for a specific group. Work to characterize sociotechnical harms
can narrow the scope .
. Value judgments for gating and limiting access
A base generative AI system is capable of many types of content making content moderation complex
. What constitutes appropriate outputs is influenced by religion cultural and personal beliefs.
What content can and should be limited filtered and gated is also vague. For example sexual content
may not be inherently unsafe to generate in some cultures but may be subject to local laws. Technical
filters may not be able to distinguish between sexual content and nonsexual nudity and may not be
able to distinguish between consensual and nonconsensual content. While most difficult without
specific use cases or context specific applications face the same challenges.
The Gradient of System Access
Once considerations are taken into account the group determining release method must choose if the
system and its components are publicly acknowledged and released. The below gradient of release
options are based on five years of publicized generative AI systems. This gradient of
options serves as a framework and does not fully capture the nuance of the many components and
details in a system release.
Figure shows the tradeoffs in considerations along the gradient as systems become more open they
better enable audits and community research but are more difficult to control for risks.
internal research only
high risk control
low auditability
limited perspectives
community research
low risk control
high auditability
broader perspectives
fully closed gradualstaged release hosted access cloudbasedAPI access downloadable fully open
PaLM Google
Gopher DeepMind
Imagen Google
MakeAVideo Meta
GPT OpenAI
Stable Diffusion Stability AI
DALLE OpenAI
Midjourney Midjourney
GPT OpenAI OPT Meta
Craiyon craiyon
BLOOM BigScience
GPTJ EleutherAI
gated to public
System
Developer
L
e
v
el o
f
A
c
c
e
s
s
Considerations
Figure Considerations and Systems Along the Gradient of System Access
Below the gradient are examples of generative systems placed according to their original release
method upon announcement for example GPT may fall under Downloadable today but was
originally released as GradualStaged.
. Fully Closed
When all aspects and components of a system are inaccessible outside the developer organization or
even closed outside a specific subsection of an organization the system is fully closed. At the furthest
end of the spectrum the systems existence is unknown outside a select group within the developer
organization even after full training. A fully closed system may or may not include some form of
public announcement that the system exists. These systems can only be researched by the developer
organization which is often a highresource organization such as the Alphabet companies Google
and DeepMind. Some publiclyknown systems such as Googles Imagen and DeepMinds
Gopher are examples. Public engagement may come from the system being deployed in a
commercial application or the public calling out biases and notable social aspects of a system from
public releases. These releases can be cherrypicked for example showing only nonhuman animals
or human silhouettes which does not give robust insight to broad capabilities or social impacts
such as biases.
. GradualStaged Release
This method refers to releasing a system in stages or gradually over a predetermined amount of
time. The time between stages is intended for investments that minimize risk such as monitoring for
malicious actor activity and conducting research on potential harms. In OpenAI stagereleased
language model GPT in four sizes by increasing parameter count over nine months while conducting
research internally and with external partners . This sparked debate among some but still
is a recommended tactic among others . In Stability AIs Stable Diffusion initially
approached a stage release by providing access to a hosted model before releasing the model weights.
However model weights were leaked days after their initial hosted release. This exemplifies
the need to inject safety protocols and prevent leakage during this approach. While there is no
standardized time frame for staged releases generally substantial sociotechnical research requires
multiple weeks months and sometimes years.
. Gated to Public Access Including Paid and Free
When providing access to a system without fully opening all components actors deciding release
method may choose to place access limitations. Above the infrastructural limitation options namely
hosting cloudbased access or fully downloadable access is the choice to make the release gated
or public. Gating system access is a selective process by group of people used to block high risk
or outofscope use cases. Limited access can make enforcing controls easier for example system
deployers withhold the right to revoke access in a gated and hosted access setting. However gating
downloadable systems is unreliable as a technical mechanism the network effect of researchers
sharing within the same circles can provide a loophole to gating. The releasing organization
cannot fully monitor whether users are sharing access through screensharing credentialsharing or
simply sending components such as model weights to unauthorized users. This does not mean this is
an ineffective guardrail as it still creates barriers to sharing the model. The deploying organization
will still be making critical decisions that contribute to concentration of power.
.. Hosted Access
System deployers may provide access to the model itself by hosting the model on their own servers
and allowing surfacelevel interfacing. Access can differ depending on the interfaces usability
especially for users with minimal or no experience with these systems. Generally users are unable
to perform tasks outside what is prescribed usually simple inputoutput probing. This method is
specific only to model access not access to other system components. Examples include Midjourney
which allows users to interact with its image generation model via Discord bot or web interface
. When optimized for usability and dialogue as seen with OpenAIs ChatGPT broader
perspectives can interact with and the model but raises misuse concern and ethical challenges .
This method can also transition to API or downloadable as seen with OpenAIs DALLE was
switching to APIbased access seven months later . While this method provides some model
access it limits external research ability.
.. CloudbasedAPI Access
Cloudbased access or access provided via application programming interface API provides more
insight and researchability into a model than Hosting but still allows for restrictive functionality.
Some APIs only allow for querying such as OpenAIs original GPT release via API . Additional
functionality can be added such as finetuning via API. Similar to Hosting this method is specific
only to model access. Nonreleased components and system information can be determined via tools
such as EleutherAIs evaluation harness used to determine GPT parameter sizes via OpenAIs
API . This method is favorable for structured access where research is possible but can still be
tracked and is unlikely to create a modified version . Cloudbased access can track users and
their activity to monitor for risky behavior. This can also better enforce safety controls such as rate
limiting. .
.. Downloadable
The main distinction between downloadable and fully open systems is the withholding of system
components such as training dataset availability. Downloadable systems can also be gated. Downloadability does not inherently imply full access to any user granted access as the size of a model can
limit who is capable of running a modal locally. Personal and standard consumer hardware is unlikely
to support large and powerful models. The infrastructure needed to run large models creates an access
barrier. In response industry and public initiatives are creating accessible infrastructure
for researchers. Downloadable models better enable robust research but are difficult to track for
potential misuse or harm. This method also eases user ability to erode or disable safety controls such
as content filters.
. Fully Open
When all aspects of the system are accessible and downloadable including all components the system
is fully open. These systems cannot be gated and by definition are fully public. For the purposes of
this framework a basic level of accessibility and documentation across components qualifies a system
release as fully open but releases may differ in documentation detail and levels of granularity. The
most prominent fully open systems were developed by organizations founded on the principle of
openness. EleutherAI is a decentralized collective also prioritizes transparency and has released all
system components as seen in their GPTJ and GPTNeo language models and the Pile
dataset . The BigScience global research community of over researchers developed the
BLOOM language model in the open . Over working groups covered aspects from dataset
creation to carbon footprint to modeling approach to optimize for a multilingual system created
transparently . While openness does enable broader research that can engage many peoples it
can also enable dangerous uses and model creation and controls can be difficult to enforce.
Trends in System Releases
We analyze release trends across prominent base generative AI systems this does not include finetuned or updated systems such as models that undergo reinforcement learning with human feedback.
These figures are based on tracking and evaluation initiatives are not exhaustive and
intend to show release trends over time.
. Timelines for Large Language Models
When examining systems by the original method of release over time trends seen in Figure show
closing and limiting language model access as more common since GPTs staged release. Language
models with fewer than six billion parameters have generally been towards the open end of the
gradient but more powerful models especially from large companies tend to be closed. This can be
due to their requiring deeper consideration and safeguards due to risk potential but Figure also
illustrates the high number of large companies able to develop and close language models.
Fully Closed
GradualStaged
Hosted
CloudbasedAPI
Downloadable
Fully Open
Figure Language Model Release Method By Parameter Count Over Time
. Timelines for All Modalities
As more generative modalities are developed from image to audio to video they face similar release
decision challenges.
Figure shows system release over this same time period. As there is no standard means to compare
capabilities across modalities all levels of system capability are placed equally. Again trends show
openness until GPTs staged released. This timeline also shows a sharp increase in the amount of
systems developed and closed after . The systems most commonly toward the open end of the
gradient are developed by smaller organizations founded with the intent to be open.
Conversely many systems from large companies
are becoming closed or have closed components.
OpenAI is the most common company to restrict but not fully close or open access. Alphabet
companies Google and DeepMind are most common among closed systems. Across modalities
large companies have steered toward closedness. Open initiatives from large companies are shown
to release a downloadable model trained on public datasets crafted by other organizations as seen
with Metas OPTB . It is unclear at this time whether movements towards openness will
pressure historically closed organizations to adjust their release strategies.
Safety Controls and Guardrails
A combination of controls and guardrails largely from the developer and deploying organizations but
also from external researchers can complement each other in order to address the above considerations
and risks. Many of these methods are pioneered and honed in research environments and outside
developer organizations. Individually no one control can serve as a panacea. While it is possible to
add controls and guardrails long after deployment these options are most effective when deployed
simultaneously with system release.
. Documentation and Transparency
Structured documentation that clearly communicates critical information about each component of
the system gives further insight to the system and can take many forms. Proposed approaches to documentation at dataset and model levels have proven successful without any enforcement mechanisms
many releases across AI companies include some form of this documentation. Datasheets for datasets
See Appendix A for logo and developer key
GPT
GPT
Turing
NLG
Jukebox
GPT
DALLE
BERT
CTRL
Gopher
GPTNeo
CPM
M
PanGu
ERNIE
.
Wu
Dao
Jurassic Hyper
CLOVA
Yuan
ERNIE
Titan
Cohere
LM
PAG
nol
XGLM
LaMDA Megatron
Turing GPT NLG
GPT NeoX
SW
Chinchilla PaLM
DALL
E
Lyrafr
Imagen OPT
Craiyon
YaLM Minerva
Midjourney BLOOM Stable
Diffusion
MakeAVideo Phenaki Audio
LM
ELMo
GPTJ
Galactica Chat
GPT
Anthropic
LM
Fully Closed
GradualStaged
Hosted
CloudbasedAPI
Downloadable
Fully Open
Figure Release Methods Over Time All Modalities
communicates aspects of datasets such as creators motivations collection process and overall
composition. Metas OPTB release included a datasheet in its appendix . Data statements
for natural language processing are another popular tool more tailored to languagebased systems
seen used by bias measurement dataset CrowSPairs . Model cards have been popular as
seen in Googles PaLM OpenAIs GPT and GPT and Runway Research and
Stability AIs Stable Diffusion . Model cards are deployed across Hugging Faces platform and
have evolved to be interactive . System cards blend datasheets and model cards and have
been used for DALLE .
. Technical Tools
Technical tools can address specific technical safety concerns but cannot be a substitute for addressing
complex societal problems. In some cases technical tools can create new social harms and should
therefore be vetted and combined with other guardrails.
.. Rate Limiting
Constricting the amount of outputs a user can generate via cloudbased access is a popular means
of preventing attacks and harmful generations. Rate limiting also helps a system perform well and
protect underlying infrastructure from being overloaded. This defensive measure can be enforced
with common strategies such as a token bucket which tracks and limits usage according to a set
number of tokens that can refresh or accumulate on a predetermined time frame. As an example
OpenAIs DALLE s public API rate limits external users . This can be adapted for users whose
applications have been cleared as safe.
.. Safety and Content Filters
Filters developed to trigger blank responses when given an unsafe input are popularly deployed across
varying levels of access. This can help block illegal and egregious content. Developers selecting
these trigger categories must make normative judgements about what input content blocks generation.
Stable Diffusions safety filter was found to primarily prevent generations with sexually explicit
content but not violence and gore which is a normative judgment about the safety of both
categories in generated images. Blocking generations for socially sensitive topics can result in entire
identity groups being blocked. Lessons from social media platform content moderation highlight
harms such as community erasure especially among marginalized groups .
.. Detection Models
While methods to detect AI generated outputs can vary and include human detection detection
models can be a helpful tool especially for less powerful generative systems. While the human
eye alone can detect outputs from less powerful systems such as Craiyon for AIgenerated images
detection models can have higher accuracy for more powerful systems. This is particularly important
when models are deployed in highstakes settings . As system output quality improves that
distinction becomes more difficult for both humans and AI detection models. Approaches to detection
can be tailored to modality such as text and include human annotation . Detection
models can also differ based on type of generation within a modality such as facial generations .
.. Hardcoding Responses
Predetermined safe outputs triggered for a given input can be hardcoded into a model interface. This
can aid legal compliance or provide standardized responses for highrisk inputs. Similar to filtering
determining trigger inputs or trigger categories requires normative judgements about what constitutes
unsafe inputs and what constitutes appropriate outputs. This can not only lead to community erasure
but also impose these normative beliefs onto users.
.. Watermarking
The concept of digital watermarking media can be transferred to AI systems to protect against model
theft protect IP and more easily identify AIgenerated outputs. Encoding a unique identifier in
generated outputs can aid in detecting media as AIgenerated and synthetic and trace the output to a
specific model. Research strives ensure these watermarks are invisible to the humaneye do not
affect output quality and tamperproof from model attacks and alterations like finetuning via
methods such as embedding noise as watermarks . Different approaches to watermarking can
be deployed for different needs from the embedding method to easily determine whether an output is
synthetic to linking watermarks to a model owners identity for authentication purposes . There
are no current prominent successful case studies as watermarking has not yet been publicly deployed
at scale for large generative systems.
.. Model Weight Encryption
Encryption can be used in order to protect model weights often to protect from model stealing and
to protect IP. This allows only an authorized user with the key to use the model. s proposed
NNLock does not change model structure so as not to adversely affect model performance.
proposed an obfuscation framework that only authorizes users with a trustworthy hardware device.
notes many existing IP protection methods are not robust to model attacks and are not suitable
for commercial purposes as they verify model ownership but not user identities.
.. Updating Adapting or Retraining models
Models can be adapted in a way that mitigates risk. Popular methods include finetuning for example
finetuning GPT on valuestargeted datasets or finetuning LaMDA on annotated data to
improve factual grounding . Another method is reinforcement learning with human feedback as
seen with InstructGPT and its opensource replication effort at CarperAI . These methods
result in new models different from their base models but often improved along a safety parameter.
. Community and Platform Efforts
Communitydriven approaches to risk mitigation leverage new and varied viewpoints. Bounty
programs from bug bounties to bias bounties can raise unforeseen safety issues and strengthen
trust in a system . Bias bounties by nature benefit from diverse perspectives. Community
moderation communitybased content flagging and naming and shaming techniques on a platform
enables users to determine and stop harmful content before it escalates. Monitoring and logging
inputs by a user on the backend helps track trends in harmful or extremist behavior.
. Organizational and Platform Policies
Organizational and platform policies can guide and enforce safe human interaction with generative
AI systems. These policies can have drawbacks they may protect from harm but also limit beneficial
uses. For example limiting access in a region under active war can prevent disinformation generation
but also general access. Internal risk policies should provide a process for what considerations must
be weighed and how to evaluate each prior to determining release options. If the system is deployed
on a platform or on a given interface a code of conduct for engaging with the platform and other
users prevents direct harm on the platform at risk of losing platform access. Mandating user accounts
on a platform helps track specific users and their activity which supports community and platform
efforts. Sharing policies that outline what can and cannot be posted on other platforms or for uses
outside of personal use prevents harmful content from spreading and inciting further harmful content.
. Legal Recourse
Legal measures such as licenses are an enforceable control when a user uses a system in a way the
deployer prohibited. The Responsible AI License RAIL places behavioral use conditions on a
model with the model owner owning the license and responsibility for pursuing enforcement if need
be . Both BigSciences BLOOM model and Runway Research and Stability AIs Stable
Diffusion use RAILs. Licenses are difficult to enforce for downloadable or fully open systems
as model behavior and uses cannot be fully monitored. Legal enforcement can also be costly in terms
of both time and financial resources. Example cases studies are examined in .
Necessary Investments for Responsible Release
Developers and researchers must listen and leverage multidisciplinary and often external expertise
especially for guardrails. Policymakers must mandate safety where possible and technically feasible
and provide resources for the underresourced. Regardless of level of access generative AI systems
are capturing reflecting and amplifying aspects of society that require multiple perspectives in
exploratory and risk control research. Since a system cannot be fully safe or unbiased for all groups
of peoples and there is no clear standard for when a system is safe for broad public release further
discourse across all affected parties is needed. Research and decisions made now will inform
considerations for increasingly powerful systems across modalities in the future making early
investment crucial.
. Accessible Interfaces and Low and Nocode Tools
In order to make generative systems accessible to the many peoples they affect means of interacting
with a system such as a model demo are needed. A clean easily usable interface that accommodates
disabilities and all levels of technical comfort significantly improves accessibility. This step towards
further openness can push a system toward the far end of the gradient with less risk control and
increased redteaming. Accessible interfaces with lowbarrier sharing can also better enable crossfield collaborations . Largescale probing can reveal flaws as seen with Metas Galactica language
model which was released with a demo. The demo was retracted within three days due to the public
naming risks such as disinformation generation . Both computer science training and low to
nocode interfaces are necessary to streamline sociotechnical research. Moral experiments show
varying approaches to ethical problems by background and culture which are urgently needed
perspectives in building evaluating and deploying new AI systems. Effective design and user
interface must be optimized for experts outside of computer science .
. Closing Resource Gaps
Resource gaps mainly between among major labs research groups and academia have widened .
In addition to the gap hindering groups from developing systems at the same level of performance it
also hinders the ability to build and run exploratory research projects. The monetary infrastructure
and sometimes skills limitations can bar especially underrepresented groups from contributing to
understanding and mitigating risks. Public sector investment at national and global levels can
start to bridge this gap. Grants from developer labs can also sponsor thirdparty research but should
have builtin mechanisms for also allowing critical research. Infrastructure grants for computer
clusters can enable smaller research groups to engage with powerful systems. Skillbuilding requires
longerterm investment.
. Technical and Practical Ethics Training
Increasing access to social scientists and the many multidisciplinary experts underrepresented in
the AI research community is insufficient. The technical barriers to evaluating and improving or
mitigating harms of AI systems can slow or hinder critical research. Conversely the lack of practical
ethics and science and technology studies STS training among technical professionals prevents
thoughtfully integrating societal guardrails from project conception throughout the development
process. Training must be implemented at early stage education academic courses and curricula in
computer science must integrate social and ethical considerations. Social sciences geared toward
examining AI systems must foster technical understanding.
. Expert Foresight
Experts in relevant disciplines should be included while relative risk is low. As generative AI systems
become a higher risk for specific applications and fields such as disinformation and medical advice
correlating experts should be tapped in and begin foundational work in mitigating that risk. The rapid
rate of AI development means substantial research should be planned in anticipation to prevent the
trend of detection and mitigation trailing capabilities advancement. For example research conducted
on radicalization risks of language models using GPT and GPT show that GPT has significantly
higher risk potential in generating extremist text . Starting this research with less powerful
systems can better inform future mitigation efforts.
. Multidisciplinary Discourse
Increasing access to social scientists and the many multidisciplinary experts underrepresented in
this discussion is insufficient. Critically actors in this space must have some incentive to engage in
frequent community discussion and be held accountable to commitments for safe releases. Googles
public position on responsible AI practices encourages inhouse risk evaluation and mitigation but
conflicts of interest can result in internal critics being unable to share or publish findings and
dismissal . These initiatives can also be formed as an industry argument for selfregulation
but ultimately lack external accountability. A new thirdparty convening body can help facilitate
this discourse. Instead of relying on existing fora conversations can take lessons from social and
abolitionist movements in how to include underrepresented and affected communities .
. Enforcement Mechanisms for Unsafe Release
Ultimately all actors involved in the release of a powerful AI system must have some incentive to
conduct releases safely. But enforcing responsible release requires a definition for what constitutes
responsible release. Responsible is distinct from safe and can emphasize meeting all possible
enforceable safety guardrails pre and postrelease. Regulation can mandate that releases include
system documentation and auditing for highrisk or highimpact releases. Updatable policies can
recommend certain risk controls and guardrails and policy bodies can better fund risk research and
development of further evaluations and controls.
Conclusion
The gradient of generative AI system release shows the complexity and tradeoffs of any one option.
Releases must balance concentration of power and AI risks in addition to considering precedent for
future releases as system capabilities increase. Developers and deployers regardless of release method
preference must engage multidisciplinary experts and the AI community to better form norms for safe
release. Existing and evolving risk controls and guardrails require developer deployer researcher
and policymaker action and can mitigate some foreseeable harms but longterm investments in
disciplines and discourse across the AI community and among affected peoples are necessary.
Acknowledgments and Disclosure of Funding
Thank you to Hugging Face for funding this research.
Thank you to Joshua Achiam Stella Biderman Miles Brundage Clmentine Fourrier Yacine Jernite
Margaret Mitchell Percy Liang and Sonja SchmerGalunder for their thoughtful feedback on earlier
versions of this paper.
Thank you to Johann Christensen for testing figures.
References
A. Abid A. Abdalla A. Abid D. Khan A. Alfozan and J. Y. Zou. Gradio Hasslefree
sharing and testing of ML models in the wild. CoRR abs. . URL http
arxiv.orgabs..
M. Alam S. Saha D. Mukhopadhyay and S. Kundu. Nnlock A lightweight authorization to
prevent ip threats of deep learning models. J. Emerg. Technol. Comput. Syst. apr .
ISSN . doi .. URL httpsdoi.org..
E. Awad S. Dsouza R. Kim J. Schulz J. Henrich A. Shariff J. F. Bonnefon and
I. Rahwan. The moral machine experiment. Nature . URL
httpswww.nature.comarticless.
N. Benaich and I. Hogarth . URL httpswww.stateof.aireportlaunch.
html.
E. M. Bender and B. Friedman. Data statements for natural language processing Toward
mitigating system bias and enabling better science. Transactions of the Association for
Computational Linguistics . doi .tacla. URL https
aclanthology.orgQ.
E. M. Bender T. Gebru A. McMillanMajor and S. Shmitchell. On the dangers of stochastic
parrots Can language models be too big? . In Proceedings of the ACM Conference on
Fairness Accountability and Transparency FAccT page New York NY USA
. Association for Computing Machinery. ISBN . doi ..
. URL httpsdoi.org...
R. Benjamin. Race after technology Abolitionist Tools for the new jim code. Polity .
Y. Benkler. Dont let industry write the rules for ai. Nature News May . URL https
www.nature.comarticlesd.
A. S. Bergman G. Abercrombie S. Spruit D. Hovy E. Dinan Y.L. Boureau and V. Rieser.
Guiding the release of safer EE conversational AI through value sensitive design. In Proceedings of the rd Annual Meeting of the Special Interest Group on Discourse and Dialogue
pages Edinburgh UK Sept. . Association for Computational Linguistics. URL
httpsaclanthology.org.sigdial..
BigScience. Bigscience rail license v. . URL httpshuggingface.cospaces
bigsciencelicense.
A. Birhane P. Kalluri D. Card W. Agnew R. Dotan and M. Bao. The values encoded
in machine learning research. In ACM Conference on Fairness Accountability and
Transparency FAccT page New York NY USA . Association for
Computing Machinery. ISBN . doi ... URL
httpsdoi.org...
S. Black S. Biderman E. Hallahan Q. Anthony L. Gao L. Golding H. He C. Leahy
K. McDonell J. Phang M. Pieler U. S. Prashanth S. Purohit L. Reynolds J. Tow B. Wang
and S. Weinbach. GPTNeoXB An opensource autoregressive language model. In
Proceedings of the ACL Workshop on Challenges Perspectives in Creating Large Language
Models . URL httpsarxiv.orgabs..
S. L. Blodgett S. Barocas H. Daum III and H. Wallach. Language technology is
power A critical survey of bias in NLP. In Proceedings of the th Annual Meeting
of the Association for Computational Linguistics pages Online July .
Association for Computational Linguistics. doi .v.aclmain.. URL
httpsaclanthology.org.aclmain..
F. Boenisch. A systematic review on model watermarking for neural networks. Frontiers
in Big Data . ISSN X. doi .fdata... URL https
www.frontiersin.orgarticles.fdata...
R. Bommasani D. A. Hudson E. Adeli R. Altman S. Arora S. von Arx M. S. Bernstein
J. Bohg A. Bosselut E. Brunskill E. Brynjolfsson S. Buch D. Card R. Castellon N. S.
Chatterji A. S. Chen K. Creel J. Q. Davis D. Demszky C. Donahue M. Doumbouya
E. Durmus S. Ermon J. Etchemendy K. Ethayarajh L. FeiFei C. Finn T. Gale L. Gillespie
K. Goel N. D. Goodman S. Grossman N. Guha T. Hashimoto P. Henderson J. Hewitt
D. E. Ho J. Hong K. Hsu J. Huang T. Icard S. Jain D. Jurafsky P. Kalluri S. Karamcheti
G. Keeling F. Khani O. Khattab P. W. Koh M. S. Krass R. Krishna R. Kuditipudi and et al.
On the opportunities and risks of foundation models. CoRR abs. . URL
httpsarxiv.orgabs..
G. Brockman M. Murati P. Welinder and OpenAI. Openai api Jun . URL https
openai.comblogopenaiapi.
M. Brundage S. Avin J. Clark H. Toner P. Eckersley B. Garfinkel A. Dafoe P. Scharre
T. Zeitzoff B. Filar H. S. Anderson H. Roff G. C. Allen J. Steinhardt C. Flynn S. .
higeartaigh S. Beard H. Belfield S. Farquhar C. Lyle R. Crootof O. Evans M. Page
J. Bryson R. Yampolskiy and D. Amodei. The malicious use of artificial intelligence
Forecasting prevention and mitigation. CoRR abs. . URL httparxiv.
orgabs..
M. Brundage S. Avin J. Wang H. Belfield G. Krueger G. Hadfield H. Khlaaf J. Yang
H. Toner R. Fong T. Maharaj P. W. Koh S. Hooker J. Leung A. Trask E. Bluemke
J. Lebensold C. OKeefe M. Koren T. Ryffel J. Rubinovitz T. Besiroglu F. Carugati
J. Clark P. Eckersley S. de Haas M. Johnson B. Laurie A. Ingerman I. Krawczuk A. Askell
R. Cammarota A. Lohn D. Krueger C. Stix P. Henderson L. Graham C. Prunkl B. Martin
E. Seger N. Zilberman S. . higeartaigh F. Kroeger G. Sastry R. Kagan A. Weller B. Tse
E. Barnes A. Dafoe P. Scharre A. HerbertVoss M. Rasser S. Sodhani C. Flynn T. K.
Gilbert L. Dyer S. Khan Y. Bengio and M. Anderljung. Toward trustworthy ai development
Mechanisms for supporting verifiable claims . URL httpsarxiv.orgabs.
.
L. Cai and Y. Zhu. The challenges of data quality and data quality assessment in the big data
era. Data Sci. J. .
A. Caliskan J. J. Bryson and A. Narayanan. Semantics derived automatically from language
corpora contain humanlike biases. Science . doi .science.
aal. URL httpswww.science.orgdoiabs.science.aal.
N. Carlini F. Tramr E. Wallace M. Jagielski A. HerbertVoss K. Lee A. Roberts T. B.
Brown D. Song . Erlingsson A. Oprea and C. Raffel. Extracting training data from large
language models. CoRR abs. . URL httpsarxiv.orgabs.
.
CarperAI. Carperai an eleutherai lab announces plans for the first opensource instructiontuned language model. Oct . URL httpscarper.ai
instructgptannouncement.
A. Chakraborty A. Mondal and A. Srivastava. Hardwareassisted intellectual property
protection of deep learning models. In Proceedings of the th ACMEDACIEEE Design
Automation Conference DAC . IEEE Press . ISBN .
A. Chowdhery S. Narang J. Devlin M. Bosma G. Mishra A. Roberts P. Barham H. W.
Chung C. Sutton S. Gehrmann P. Schuh K. Shi S. Tsvyashchenko J. Maynez A. Rao
P. Barnes Y. Tay N. Shazeer V. Prabhakaran E. Reif N. Du B. Hutchinson R. Pope
J. Bradbury J. Austin M. Isard G. GurAri P. Yin T. Duke A. Levskaya S. Ghemawat
S. Dev H. Michalewski X. Garcia V. Misra K. Robinson L. Fedus D. Zhou D. Ippolito
D. Luan H. Lim B. Zoph A. Spiridonov R. Sepassi D. Dohan S. Agrawal M. Omernick
A. M. Dai T. S. Pillai M. Pellat A. Lewkowycz E. Moreira R. Child O. Polozov K. Lee
Z. Zhou X. Wang B. Saeta M. Diaz O. Firat M. Catasta J. Wei K. MeierHellstern D. Eck
J. Dean S. Petrov and N. Fiedel. Palm Scaling language modeling with pathways .
URL httpsarxiv.orgabs..
D. Contractor D. McDuff J. K. Haines J. Lee C. Hines and B. J. Hecht. Behavioral use
licensing for responsible AI. CoRR abs. . URL httpsarxiv.orgabs
..
A. Crisan M. Drouhard J. Vig and N. Rajani. Interactive model cards A humancentered
approach to model documentation. In ACM Conference on Fairness Accountability and
Transparency. ACM jun . doi ... URL httpsdoi.org
.F..
L. Dabbish C. Stuart J. Tsay and J. Herbsleb. Social coding in github Transparency and
collaboration in an open software repository. In Proceedings of the ACM Conference on
Computer Supported Cooperative Work CSCW page New York NY USA
. Association for Computing Machinery. ISBN . doi ..
. URL httpsdoi.org...
P. Dave and J. Dastin. Google told its scientists to strike a positive tone in ai research documents. Reuters . URL https
www.reuters.comarticleusalphabetgoogleresearchfocus
googletolditsscientiststostrikeapositivetoneinairesearchdocumentsidUSKBNXCB Y. Dou M. Forbes R. KoncelKedziorski N. A. Smith and Y. Choi. Scarecrow A framework
for scrutinizing machine text. CoRR abs. . URL httpsarxiv.org
abs..
C. Ebell R. BaezaYates R. Benjamins H. Cai M. Coeckelbergh T. Duarte M. Hickok
A. Jacquet A. Kim J. Krijger J. Macintyre P. Madhamshettiwar L. Maffeo J. Matthews
L. Medsker P. Smith and S. Thais. Towards intellectual freedom in an ai ethics global
community. AI and Ethics . doi .s.
A. Fan S. Ilic T. Wolf and M. Gall editors. Proceedings of BigScience Episode
Workshop on Challenges Perspectives in Creating Large Language Models virtualDublin
May . Association for Computational Linguistics. URL httpsaclanthology.org
.bigscience..
L. Gao. On the sizes of openai api models May . URL httpsblog.eleuther.ai
gptmodelsizes.
L. Gao S. Biderman S. Black L. Golding T. Hoppe C. Foster J. Phang H. He A. Thite
N. Nabeshima S. Presser and C. Leahy. The pile An gb dataset of diverse text for language
modeling. CoRR abs. . URL httpsarxiv.orgabs..
T. Gebru J. Morgenstern B. Vecchione J. W. Vaughan H. M. Wallach H. D. III and
K. Crawford. Datasheets for datasets. CoRR abs. . URL httparxiv.
orgabs..
S. Gehrmann H. Strobelt and A. Rush. GLTR Statistical detection and visualization of
generated text. In Proceedings of the th Annual Meeting of the Association for Computational
Linguistics System Demonstrations pages Florence Italy July . Association
for Computational Linguistics. doi .vP.
T. Gillespie. Custodians of the internet Platforms content moderation and the hidden
decisions that shape social media. Yale University Press .
J. A. Goldstein G. Sastry M. Musser R. DiResta M. Gentzel and K. Sedova. Generative language models and automated influence operations Emerging threats and potential mitigations
. URL httpsarxiv.orgabs..
D. Gragnaniello F. Marra and L. Verdoliva. Detection of AIGenerated Synthetic Faces pages
. Springer International Publishing Cham . ISBN . doi .
. URL httpsdoi.org..
O. L. Haimson D. Delmonaco P. Nie and A. Wegner. Disproportionate removals and differing
content moderation experiences for conservative transgender and black social media users
Marginalization and moderation gray areas. Proc. ACM Hum.Comput. Interact. CSCW
oct . doi .. URL httpsdoi.org..
W. D. Heaven. Why metas latest large language model survived only three days online Nov . URL httpswww.technologyreview.com
metalargelanguagemodelaionlysurvivedthreedaysgptscience.
D. Hovy and S. L. Spruit. The social impact of natural language processing. In Proceedings
of the th Annual Meeting of the Association for Computational Linguistics Volume
Short Papers pages Berlin Germany Aug. . Association for Computational
Linguistics. doi .vP. URL httpsaclanthology.orgP.
G. Jawahar M. AbdulMageed and L. V. S. Lakshmanan. Automatic detection of machine
generated text A critical survey. CoRR abs. . URL httpsarxiv.org
abs..
P. Kalluri. Dont ask if artificial intelligence is good or fair ask how it shifts power. Nature
News July . URL httpswww.nature.comarticlesd.
N. Kaloudi and J. Li. The aibased cyber threat landscape A survey. ACM Comput. Surv.
feb . ISSN . doi .. URL httpsdoi.org.
.
H. Khlaaf P. Mishkin J. Achiam G. Krueger and M. Brundage. A hazard analysis framework
for code synthesis large language models . URL httpsarxiv.orgabs.
.
J. Kirchenbauer J. Geiping Y. Wen J. Katz I. Miers and T. Goldstein. A watermark for large
language models . URL httpsarxiv.orgabs..
S. Kreps R. M. McCain and M. Brundage. All the news thats fit to fabricate Aigenerated
text as a tool of media misinformation. Journal of Experimental Political Science
. doi .XPS...
P. Liang R. Bommasani K. Creel and R. Reich. The time is now to develop community
norms for the release of foundation models . URL httpshai.stanford.edu
newstimenowdevelopcommunitynormsreleasefoundationmodels.
P. Liang R. Bommasani T. Lee D. Tsipras D. Soylu M. Yasunaga Y. Zhang D. Narayanan
Y. Wu A. Kumar B. Newman B. Yuan B. Yan C. Zhang C. Cosgrove C. D. Manning
C. R D. AcostaNavas D. A. Hudson E. Zelikman E. Durmus F. Ladhak F. Rong
H. Ren H. Yao J. Wang K. Santhanam L. Orr L. Zheng M. Yuksekgonul M. Suzgun
N. Kim N. Guha N. Chatterji O. Khattab P. Henderson Q. Huang R. Chi S. M. Xie
S. Santurkar S. Ganguli T. Hashimoto T. Icard T. Zhang V. Chaudhary W. Wang X. Li
Y. Mai Y. Zhang and Y. Koreeda. Holistic evaluation of language models . URL
httpsarxiv.orgabs..
T. Liao. Foundation model tracker. httpsfoundationmodeltracker.com .
Z. C. Lipton. Openai trains language model mass hysteria ensues Feb
. URL httpswww.approximatelycorrect.com
openaitrainslanguagemodelmasshysteriaensues.
T. Mandel J. Best R. H. Tanaka H. Temple C. Haili K. Schlectinger and R. Szeto. Lets
keep it safe Designing user interfaces that allow everyone to contribute to ai safety .
URL httpsarxiv.orgabs..
K. McGuffie and A. Newhouse. The radicalization risks of GPT and advanced neural
language models. CoRR abs. . URL httpsarxiv.orgabs.
.
J. Michael A. Holtzman A. Parrish A. Mueller A. Wang A. Chen D. Madaan N. Nangia
R. Y. Pang J. Phang and S. R. Bowman. What do nlp researchers believe? results of the nlp
community metasurvey . URL httpsarxiv.orgabs..
Microsoft. Microsoft turing academic program mstap Oct .
URL httpswww.microsoft.comenusresearchcollaboration
microsoftturingacademicprogram.
Midjourney. Quick start guide . URL httpsmidjourney.gitbook.iodocs.
P. Mishkin L. Ahmad M. Brundage G. Krueger and G. Sastry. Dalle preview risks
and limitations. . URL httpsgithub.comopenaidallepreviewblob
mainsystemcard.mdhttpsgithub.comopenaidallepreviewblob
mainsystemcard.md.
M. Mitchell S. Wu A. Zaldivar P. Barnes L. Vasserman B. Hutchinson E. Spitzer I. D.
Raji and T. Gebru. Model cards for model reporting. CoRR abs. . URL
httparxiv.orgabs..
S. Mohamed M. Png and W. Isaac. Decolonial AI decolonial theory as sociotechnical
foresight in artificial intelligence. CoRR abs. . URL httpsarxiv.org
abs..
E. Mostaque. Stable diffusion public release Sep . URL httpsstability.ai
blogstablediffusionpublicrelease.
NAIRRTF. The national artificial intelligence research resource task force nairrtf . URL
httpswww.ai.govnairrtf.
N. Nangia C. Vania R. Bhalerao and S. R. Bowman. Crowspairs A challenge dataset for
measuring social biases in masked language models. CoRR abs. . URL
httpsarxiv.orgabs..
T. T. Nguyen C. M. Nguyen D. T. Nguyen D. T. Nguyen and S. Nahavandi. Deep learning
for deepfakes creation and detection. CoRR abs. . URL httparxiv.
orgabs..
S. U. Noble. Algorithms of oppression how search engines reinforce racism. New York
University Press .
OpenAI. Gpt model card Nov . URL httpsgithub.comopenaigptblob
mastermodelcard.md.
OpenAI. Gpt model card Sep . URL httpsgithub.comopenaigptblob
mastermodelcard.md.
OpenAI. Dalle api now available in public beta Nov . URL httpsopenai.com
blogdalleapinowavailableinpublicbeta.
OpenAI. Chatgpt Optimizing language models for dialogue Nov . URL https
openai.comblogchatgpt.
OpenAI. Whats the rate limit for the dalle api? how can i request
an increase? . URL httpshelp.openai.comenarticles
whatstheratelimitforthedalleapihowcanirequestanincrease.
L. Ouyang J. Wu X. Jiang D. Almeida C. L. Wainwright P. Mishkin C. Zhang S. Agarwal
K. Slama A. Ray J. Schulman J. Hilton F. Kelton L. Miller M. Simens A. Askell P. Welinder P. Christiano J. Leike and R. Lowe. Training language models to follow instructions with
human feedback . URL httpsarxiv.orgabs..
A. Ovadya and J. Whittlestone. Reducing malicious use of synthetic media research Considerations and potential release practices for machine learning. CoRR abs. .
URL httparxiv.orgabs..
C. Procope A. Cheema D. Adkins B. Alsallakh N. Green E. McReynolds
G. Pehl E. Wang and P. Zvyagina. Systemlevel transparency of machine
learning. . URL httpsai.facebook.comresearchpublications
systemleveltransparencyofmachinelearning.
J. W. Rae S. Borgeaud T. Cai K. Millican J. Hoffmann H. F. Song J. Aslanides S. Henderson R. Ring S. Young E. Rutherford T. Hennigan J. Menick A. Cassirer R. Powell
G. van den Driessche L. A. Hendricks M. Rauh P. Huang A. Glaese J. Welbl S. Dathathri
S. Huang J. Uesato J. Mellor I. Higgins A. Creswell N. McAleese A. Wu E. Elsen
S. M. Jayakumar E. Buchatskaya D. Budden E. Sutherland K. Simonyan M. Paganini
L. Sifre L. Martens X. L. Li A. Kuncoro A. Nematzadeh E. Gribovskaya D. Donato
A. Lazaridou A. Mensch J. Lespiau M. Tsimpoukelli N. Grigorev D. Fritz T. Sottiaux
M. Pajarskas T. Pohlen Z. Gong D. Toyama C. de Masson dAutume Y. Li T. Terzi V. Mikulik I. Babuschkin A. Clark D. de Las Casas A. Guy C. Jones J. Bradbury M. Johnson
B. A. Hechtman L. Weidinger I. Gabriel W. S. Isaac E. Lockhart S. Osindero L. Rimell
C. Dyer O. Vinyals K. Ayoub J. Stanway L. Bennett D. Hassabis K. Kavukcuoglu and
G. Irving. Scaling language models Methods analysis insights from training gopher. CoRR
abs. . URL httpsarxiv.orgabs..
I. D. Raji A. Smart R. N. White M. Mitchell T. Gebru B. Hutchinson J. SmithLoud
D. Theron and P. Barnes. Closing the ai accountability gap Defining an endtoend framework for internal algorithmic auditing. In Proceedings of the Conference on Fairness
Accountability and Transparency FAT page New York NY USA . Association for Computing Machinery. ISBN . doi ...
URL httpsdoi.org...
I. D. Raji E. M. Bender A. Paullada E. Denton and A. Hanna. AI and the everything in the
whole wide world benchmark. CoRR abs. . URL httpsarxiv.org
abs..
I. D. Raji P. Xu C. Honigsberg and D. Ho. Outsider oversight Designing a third party
audit ecosystem for ai governance. In Proceedings of the AAAIACM Conference on
AI Ethics and Society AIES page New York NY USA . Association
for Computing Machinery. ISBN . doi ... URL
httpsdoi.org...
A. Ramesh P. Dhariwal A. Nichol C. Chu and M. Chen. Hierarchical textconditional image
generation with clip latents . URL httpsarxiv.orgabs..
J. Rando D. Paleka D. Lindner L. Heim and F. Tramr. Redteaming the stable diffusion
safety filter . URL httpsarxiv.orgabs..
R. Reich. Now ai can write students essays for them will everyone become a
cheat? Nov . URL httpswww.theguardian.comcommentisfreenov
aistudentsessayscheatteachersplagiarismtech.
R. Rombach and P. Esser. Stablediffusion license Aug . URL httpsgithub.com
CompVisstablediffusionblobmainLICENSE.
R. Rombach A. Blattmann D. Lorenz P. Esser and B. Ommer. Highresolution image
synthesis with latent diffusion models. In Proceedings of the IEEECVF Conference on
Computer Vision and Pattern Recognition CVPR pages June .
J. Rubinovitz. Bias bounty programs as a method of combatting
bias in ai . URL httpsrubinovitz.com
biasbountyprogramsasamethodofcombatting.
C. Saharia W. Chan S. Saxena L. Li J. Whang E. Denton S. K. S. Ghasemipour B. K. Ayan
S. S. Mahdavi R. G. Lopes T. Salimans J. Ho D. J. Fleet and M. Norouzi. Photorealistic
texttoimage diffusion models with deep language understanding . URL https
arxiv.orgabs..
G. Sastry. Beyond release vs. not release . URL httpscrfm.stanford.edu
commentarysastry.html.
R. Shelby S. Rismani K. Henne A. Moon N. Rostamzadeh P. Nicholas N. Yilla J. Gallegos
A. Smart E. Garcia and G. Virk. Sociotechnical harms Scoping a taxonomy for harm
reduction . URL httpsarxiv.orgabs..
T. Shevlane. Structured access an emerging paradigm for safe ai deployment . URL
httpsarxiv.orgabs..
T. Shevlane and A. Dafoe. The offensedefense balance of scientific knowledge Does
publishing AI research reduce misuse? CoRR abs. . URL httparxiv.
orgabs..
T. Simonite. It began as an aifueled dungeon game. it got much darker May . URL
httpswww.wired.comstoryaifueleddungeongamegotmuchdarker.
U. Singer A. Polyak T. Hayes X. Yin J. An S. Zhang Q. Hu H. Yang O. Ashual O. Gafni
D. Parikh S. Gupta and Y. Taigman. Makeavideo Texttovideo generation without
textvideo data . URL httpsarxiv.orgabs..
I. Solaiman and C. Dennison. Process for adapting language models to society palms with
valuestargeted datasets .
I. Solaiman M. Brundage J. Clark A. Askell A. HerbertVoss J. Wu A. Radford and
J. Wang. Release strategies and the social impacts of language models. CoRR abs.
. URL httparxiv.orgabs..
P. Staff. Publication norms for responsible ai Dec . URL httpspartnershiponai.
orgworkstreampublicationnormsforresponsibleai.
L. Stark D. Greene and A. Hoffmann. Critical Perspectives on Governance Mechanisms
for AIML Systems pages . . ISBN . doi .
.
Z. Talat A. Nvol S. Biderman M. Clinciu M. Dey S. Longpre S. Luccioni M. Masoud
M. Mitchell D. Radev S. Sharma A. Subramonian J. Tae S. Tan D. Tunuguntla and
O. Van Der Wal. You reap what you sow On the challenges of bias evaluation under
multilingual settings. In Proceedings of BigScience Episode Workshop on Challenges
Perspectives in Creating Large Language Models pages virtualDublin May .
Association for Computational Linguistics. doi .v.bigscience.. URL
httpsaclanthology.org.bigscience..
R. Thoppilan D. De Freitas J. Hall N. Shazeer A. Kulshreshtha H.T. Cheng A. Jin T. Bos
L. Baker Y. Du Y. Li H. Lee H. S. Zheng A. Ghafouri M. Menegali Y. Huang M. Krikun
D. Lepikhin J. Qin D. Chen Y. Xu Z. Chen A. Roberts M. Bosma V. Zhao Y. Zhou C.C.
Chang I. Krivokon W. Rusch M. Pickett P. Srinivasan L. Man K. MeierHellstern M. R.
Morris T. Doshi R. D. Santos T. Duke J. Soraker B. Zevenbergen V. Prabhakaran M. Diaz
B. Hutchinson K. Olson A. Molina E. HoffmanJohn J. Lee L. Aroyo R. Rajakumar
A. Butryna M. Lamm V. Kuzmina J. Fenton A. Cohen R. Bernstein R. Kurzweil B. AgueraArcas C. Cui M. Croak E. Chi and Q. Le. Lamda Language models for dialog applications
. URL httpsarxiv.orgabs..
M. Trajtenberg. Ai as the next gpt a politicaleconomy perspective. Working Paper National Bureau of Economic Research January . URL httpwww.nber.orgpapers
w.
L. R. Varshney N. S. Keskar and R. Socher. Limits of detecting text generated by largescale
language models. CoRR abs. . URL httpsarxiv.orgabs.
.
A. Vee. Automated trolling The case of gptchan when artificial intelligence is as easy as
writing. Interfaces Essays and Reviews in Computing and Culture Vol. Charles Babbage
Institute University of Minnesota pages .
B. Wang. MeshTransformerJAX ModelParallel Implementation of Transformer Language Model with JAX. httpsgithub.comkingoflolzmeshtransformerjax
May .
L. Weidinger J. Mellor M. Rauh C. Griffin J. Uesato P.S. Huang M. Cheng M. Glaese
B. Balle A. Kasirzadeh Z. Kenton S. Brown W. Hawkins T. Stepleton C. Biles A. Birhane
J. Haas L. Rimell L. A. Hendricks W. Isaac S. Legassick G. Irving and I. Gabriel. Ethical
and social risks of harm from language models . URL httpsarxiv.orgabs
..
M. Whittaker. The steep cost of capture. Interactions nov . ISSN .
doi .. URL httpsdoi.org..
J. Whittlestone and A. Ovadya. The tension between openness and prudence in AI research.
CoRR abs. . URL httparxiv.orgabs..
D. G. Widder D. Nafus L. Dabbish and J. Herbsleb. Limits and possibilities for ethical ai
in open source A study of deepfakes. In ACM Conference on Fairness Accountability
and Transparency FAccT page New York NY USA . Association
for Computing Machinery. ISBN . doi ... URL
httpsdoi.org...
B. Workshop T. L. Scao A. Fan C. Akiki E. Pavlick S. Ilic D. Hesslow R. Castagn
A. S. Luccioni F. Yvon M. Gall J. Tow A. M. Rush S. Biderman A. Webson P. S.
Ammanamanchi T. Wang B. Sagot N. Muennighoff A. V. del Moral O. Ruwase R. Bawden
S. Bekman A. McMillanMajor I. Beltagy H. Nguyen L. Saulnier S. Tan P. O. Suarez
V. Sanh H. Laurenon Y. Jernite J. Launay M. Mitchell C. Raffel A. Gokaslan A. Simhi
A. Soroa A. F. Aji A. Alfassy A. Rogers A. K. Nitzav C. Xu C. Mou C. Emezue C. Klamm
C. Leong D. van Strien D. I. Adelani D. Radev E. G. Ponferrada E. Levkovizh E. Kim
E. B. Natan F. De Toni G. Dupont G. Kruszewski G. Pistilli H. Elsahar H. Benyamina
H. Tran I. Yu I. Abdulmumin I. Johnson I. GonzalezDios J. de la Rosa J. Chim J. Dodge
J. Zhu J. Chang J. Frohberg J. Tobing J. Bhattacharjee K. Almubarak K. Chen K. Lo
L. Von Werra L. Weber L. Phan L. B. allal L. Tanguy M. Dey M. R. Muoz M. Masoud
M. Grandury M. ako M. Huang M. Coavoux M. Singh M. T.J. Jiang M. C. Vu
M. A. Jauhar M. Ghaleb N. Subramani N. Kassner N. Khamis O. Nguyen O. Espejel
O. de Gibert P. Villegas P. Henderson P. Colombo P. Amuok Q. Lhoest R. Harliman
R. Bommasani R. L. Lpez R. Ribeiro S. Osei S. Pyysalo S. Nagel S. Bose S. H.
Muhammad S. Sharma S. Longpre S. Nikpoor S. Silberberg S. Pai S. Zink T. T. Torrent
T. Schick T. Thrush V. Danchev V. Nikoulina V. Laippala V. Lepercq V. Prabhu Z. Alyafeai
Z. Talat A. Raja B. Heinzerling C. Si D. E. Tasar E. Salesky S. J. Mielke W. Y. Lee
A. Sharma A. Santilli A. Chaffin A. Stiegler D. Datta E. Szczechla G. Chhablani H. Wang
H. Pandey H. Strobelt J. A. Fries J. Rozen L. Gao L. Sutawika M. S. Bari M. S. Alshaibani M. Manica N. Nayak R. Teehan S. Albanie S. Shen S. BenDavid S. H. Bach
T. Kim T. Bers T. Fevry T. Neeraj U. Thakker V. Raunak X. Tang Z.X. Yong Z. Sun
S. Brody Y. Uri H. Tojarieh A. Roberts H. W. Chung J. Tae J. Phang O. Press C. Li
D. Narayanan H. Bourfoune J. Casper J. Rasley M. Ryabinin M. Mishra M. Zhang
M. Shoeybi M. Peyrounette N. Patry N. Tazi O. Sanseviero P. von Platen P. Cornette P. F.
Lavalle R. Lacroix S. Rajbhandari S. Gandhi S. Smith S. Requena S. Patil T. Dettmers
A. Baruwa A. Singh A. Cheveleva A.L. Ligozat A. Subramonian A. Nvol C. Lovering
D. Garrette D. Tunuguntla E. Reiter E. Taktasheva E. Voloshina E. Bogdanov G. I. Winata
H. Schoelkopf J.C. Kalo J. Novikova J. Z. Forde J. Clive J. Kasai K. Kawamura L. Hazan
M. Carpuat M. Clinciu N. Kim N. Cheng O. Serikov O. Antverg O. van der Wal R. Zhang
R. Zhang S. Gehrmann S. Mirkin S. Pais T. Shavrina T. Scialom T. Yun T. Limisiewicz
V. Rieser V. Protasov V. Mikhailov Y. Pruksachatkun Y. Belinkov Z. Bamberger Z. Kasner
A. Rueda A. Pestana A. Feizpour A. Khan A. Faranak A. Santos A. Hevia A. Unldreaj
A. Aghagol A. Abdollahi A. Tammour A. HajiHosseini B. Behroozi B. Ajibade B. Saxena
C. M. Ferrandis D. Contractor D. Lansky D. David D. Kiela D. A. Nguyen E. Tan E. Baylor
E. Ozoani F. Mirza F. Ononiwu H. Rezanejad H. Jones I. Bhattacharya I. Solaiman
I. Sedenko I. Nejadgholi J. Passmore J. Seltzer J. B. Sanz L. Dutra M. Samagaio M. Elbadri
M. Mieskes M. Gerchick M. Akinlolu M. McKenna M. Qiu M. Ghauri M. Burynok
N. Abrar N. Rajani N. Elkott N. Fahmy O. Samuel R. An R. Kromann R. Hao S. Alizadeh
S. Shubber S. Wang S. Roy S. Viguier T. Le T. Oyebade T. Le Y. Yang Z. Nguyen A. R.
Kashyap A. Palasciano A. Callahan A. Shukla A. MirandaEscalada A. Singh B. Beilharz
B. Wang C. Brito C. Zhou C. Jain C. Xu C. Fourrier D. L. Perin D. Molano D. Yu
E. Manjavacas F. Barth F. Fuhrimann G. Altay G. Bayrak G. Burns H. U. Vrabec I. Bello
I. Dash J. Kang J. Giorgi J. Golde J. D. Posada K. R. Sivaraman L. Bulchandani L. Liu
L. Shinzato M. H. de Bykhovetz M. Takeuchi M. Pmies M. A. Castillo M. Nezhurina
M. Snger M. Samwald M. Cullan M. Weinberg M. De Wolf M. Mihaljcic M. Liu
M. Freidank M. Kang N. Seelam N. Dahlberg N. M. Broad N. Muellner P. Fung P. Haller
R. Chandrasekhar R. Eisenberg R. Martin R. Canalli R. Su R. Su S. Cahyawijaya S. Garda
S. S. Deshmukh S. Mishra S. Kiblawi S. Ott S. Sangaroonsiri S. Kumar S. Schweter
S. Bharati T. Laud T. Gigant T. Kainuma W. Kusa Y. Labrak Y. S. Bajaj Y. Venkatraman
Y. Xu Y. Xu Y. Xu Z. Tan Z. Xie Z. Ye M. Bras Y. Belkada and T. Wolf. Bloom A
bparameter openaccess multilingual language model . URL httpsarxiv.org
abs..
M. Xue C. He J. Wang and W. Liu. DNN intellectual property protection Taxonomy
methods attack resistance and evaluations. CoRR abs. . URL https
arxiv.orgabs..
J. Zhang Z. Gu J. Jang H. Wu M. P. Stoecklin H. Huang and I. Molloy. Protecting
intellectual property of deep neural networks with watermarking. In Proceedings of the
on Asia Conference on Computer and Communications Security ASIACCS page
New York NY USA . Association for Computing Machinery. ISBN .
doi ... URL httpsdoi.org...
K. A. Zhang L. Xu A. CuestaInfante and K. Veeramachaneni. Robust invisible video
watermarking with attention. CoRR abs. . URL httparxiv.orgabs
..
S. Zhang S. Roller N. Goyal M. Artetxe M. Chen S. Chen C. Dewan M. Diab X. Li
X. V. Lin T. Mihaylov M. Ott S. Shleifer K. Shuster D. Simig P. S. Koura A. Sridhar
T. Wang and L. Zettlemoyer. Opt Open pretrained transformer language models . URL
httpsarxiv.orgabs..
A Developer and Logo Key
AI Labs
AI Sweden
Alibaba
Allen Institute for AI
Anthropic
Baidu
Beijing Academy of AI
BigScience
Cohere
Craiyon
DeepMind
EleutherAI
Google
Huawei
Inspur
LightOn
Meta
Microsoft
Midjourney
Naver
NVIDIA
OpenAI
Salesforce
Stability AI
Tsinghua University
Yandex
Developer Logo Developer Logo Developer Logo
|
|
write a youtube script | arXiv.v eess.SY Jul
Distribution System Reconfiguration to Mitigate
Load Altering Attacks via Stackelberg Games
Sajjad Maleki Student Member IEEE Subhash Lakshminarayana Senior Member IEEE Charalambos
Konstantinou Senior Member IEEE and E. Veronica Belmega Senior Member IEEE
AbstractThe integration of IoTcontrollable devices in power
systems such as smart electric vehicle charging stations heat
pumps etc. despite their apparent benefits raises novel cybersecurity concerns. These vulnerabilities in these devices can
be leveraged to launch loadaltering attacks LAAs that ca
n
potentially compromise power system safety. In this paper we
analyze the impact of LAAs on the voltage profile of distribution
systems. We derive closedform expressions to quantify the
attack impact. Using the insights derived from this analysis we
propose a method to mitigate LAAs based on reconfiguring the
distribution system as a reactive defense approach. We stud
y
optimal defense strategies using a noncooperative sequential
game theory approach that is robust to LAAs. The proposed
solution takes the potential errors in the attack localization into
account. Our results show that attacks launched on the deepest
nodes in the distribution network result in the highest detrimental
impact on the grid voltage profile. Furthermore the propose
d
gametheoretic strategy successfully mitigates the effect of the
attack while ensuring minimum system reconfiguration.
Index TermsDistribution system Cybersecurity Loadaltering attack LAA Reconfiguration Stackelberg game. NOMENCLATURE
Parameters p
p Active ZP load coefficients
q
q Reactive ZP load coefficients
p p p Active ZIP load coefficients
q
q
q Reactive ZIP load coefficients
I Identity matrix b
pre
ij Component of the adjacency matrix before reconfiguration in i
th row and
j
th column
M Disjunctive parameter N Number of buses of the distribution system pd qd Nominal active and reactive power demands by a
single attacked device
p
li
q
li Active and reactive load demands in bus
i
rij xij Resistance and reactance of the line from bus
i to
j
vnom Nominal voltage
Sets
S. Maleki is with the School of Engineering University of Warwick
CVAL UK and ETIS UMR CY Cergy Paris Universite ENSEA
CNRS F Cergy France. S. Lakshimnarayana Corresponding author
is with the School of Engineering University of Warwick. C.Konstantinou
is with the CEMSE Division King Abdullah University of Science and
Technology KAUST. E. V. Belmega is with Univ. Gustave Eiffel CNRS
LIGM F MarnelaVallee France and ETIS UMR CY Cergy
Paris Universite ENSEA CNRS F Cergy France. Emails sajjad.malekiwarwick.ac.uk subhash.lakshminarayanawarwick.ac.uk charalambos.konstantinoukaust.edu.sa veronica.belmegaesiee.fr.
This work has been supported in part by PhD Cofund WALLEE project
between the University of Warwick UK and CY Cergy Paris University
France and in part by King Abdullah University of Science and Technology
KAUST under Award No. RFSOFP. The work is the extension
of a prior work accepted at IEEE PES General Meeting
.
D
i Set of buses engaged in the unique path connecting
the bus
i to the root bus
L L
s Set of lines without switches
L
s Set of lines with switches
N N
f Subset of buses which are not substations
N
a Subset of buses under attack
N
f Subset of buses which are substation
Variables vji Auxiliary voltage variable for MILP
Coefficients matrices for the proposed closedform
expressions
i Parent bus of the bus
i
Probability of the bus
to be under attack
U Vector of the square of voltages of the system while
no LAA
U
A Vector of square of voltages of the system under
LAA
a
r
a
Players actions in equilibrium
B Adjacency matrix of the distribution network after
reconfiguration
bij Component of
B in
i
th row and
j
th column
c
n
Number of attacked devices in a critical attack
in bus
n
I
ii Current flowing through the branch
i
i
L
natt The bus under attack pa
q
a Active and reactive powers raised by LAA in bus i
p
i
p
i Active and reactive load demands at rated voltage
in bus
i
p
fi
q
fi Active and reactive powers flowing from substation
bus
i
pij qij Active and reactive power flows from bus
i to
j
r
a
Followers best response to attackers strategy
s
i Apparent power in bus
i
s
zip i
s
zpi Apparent power in bus
i with ZIP and ZP load
models
v
i Voltage in bus
i
I. INTRODUCTION
InternetofThings IoT enabled devices offer enhanced
enduser convenience improved efficiency and flexibility t
o
power systems for load peak management which has driven a
notable surge in their adoption. However beyond their evident
benefits these devices also present potential vulnerabilities
serving as entry points for cyber attackers to exploit and
compromise the security of power systems. Specifically loadaltering attacks LAA in power networks with high IoTenabled device penetration pose a significant cybersecurity
threat .
The concept of LAAs was first introduced in in which
adversaries turn a group of IoTcontrollable electrical loads
into bots and turn them onoff simultaneously to harm the
stability of the system. The manipulation of loads disrupts
the balance between power generation and demand leading
to frequency instabilities in transmission networks . In
distribution networks LAA can result in elevated line flows
causing higher voltage drops leading to voltage constraint violation . Furthermore the data required to launch successful
LAAs can be obtained from publicly available information .
A. Literature Survey
LAAs have gained significant interest over the last few
years. The majority of existing works focus on the transmission system and the associated frequency control loops. We
divide these works into two groups attack impact analysis
and viability and attack mitigation.
Attack Impact Analysis and Viability Researchers in
and investigated the impact of LAAs on transmission
systems with varying levels of inertia. These studies identified
several effects encompassing line failure frequency deviation
disruption in grid restarting tieline failure and increased
operational costs. These works did not consider the inherent
protection features embedded in power grids such as N
scheduling etc. The research in examined the effects of
LAA under a more realistic setting consisting of protection and
loadshedding schemes and showed that LAA can still cause
outages and islands. The authors of and examine the
feasibility and impact of LAAs under low demand and low
inertia conditions experienced during the COVID pandemic
respectively and show that under these conditions it can be
easier to launch LAAs that lead to unsafe grid conditions.
A study on impacts of attacks on IoTenabled devices on
interconnected transmission and distribution TD systems
was carried out in .
Reference proposed the socalled dynamic LAA
DLAA in which the adversary toggles the compromised load
devices on and off continually guided by a feedback control
loop in response to the systems frequency fluctuations. An
analytical framework was introduced in to analyze the
impact of LAA on transmission systems and identify the nodes
from which an attacker can launch the most effective attacks
using the theory of secondorder dynamical systems. In a
rareevent sampling algorithm was proposed that uncovers the
spatial and temporal distribution of impactful DLAAs while
taking N security constraint into account.
Attack Mitigation Another stream of research investigates
the mitigation of LAAs. The existing mitigation methods can
be categorized into i offline or ii online methods.
Offline methods Offline defenses try to install preventive
measures to stop the detrimental impact of LAAs. For instance
reference proposes algorithms to determine the operating
points for generators in a way to prevent line overloads caused
by potential botnettype attacks against IoT load devices. In
the authors proposed the optimal placement of energy
storage units to mitigate the effect of LAAs. Reference
presented a mitigation framework based on securing the most
critical loads that can prevent the destabilizing effects of LAA.
This method finds the minimum magnitude of loads needed
to be protected in order to guarantee frequency stability in the
event of DLAAs.
While the works above focus on transmission systems
proposed a mitigation approach tailored specifically for
distribution networks. Their research focuses on identifying
optimal locations for deploying soft open points SOPs and
refining their operation to mitigate the effects of attacks on
voltage deviations from nominal value.
Online methods Despite the effectiveness of the offline
methods these measures may be too costly as the preventive
features must be enabled irrespective of whether an attack
occurs or not e.g. uneconomic generator operating points
to cover for LAAs. Online methods on the other hand
involve determining defensive actions to counter the effects
of LAAs once the attack is launched via reactive measures.
In a cyberresilient economic dispatch method has been
introduced to mitigate LAAs based on altering the frequency
droop control parameters of inverterbased resources to counter
the destabilizing effects of LAAs.
While EVs can be a source of security threat to target
the power grid the work in build a defensive
algorithm against LAA using PEVs. In this framework PEVs
are designed as feedback controllers that can mitigate the
impacts of LAA based on H and H norms. To analyze
the manoeuvres of a strategic attacker initiating DLAA
introduces a multistage game approach. In this game the
defensive actions involve load shedding and the ultimate
objective is to achieve a strategic balance between DLAA
and the necessary amount of load shedding reaching a Nash
equilibrium NE.
B. Contributions
In this paper we start with a rigorous investigation of the
impact of LAAs on distribution networks. Then we introduce
a Stackelberg noncooperative game to analyze the interaction
between the defender operator and the attacker and propose
an LAA mitigation technique in distribution networks based
on network reconfiguration via switching different lines onoff.
Finally we show the effectiveness of our proposed mitigation
method by examining it on two standard test cases the and
bus grids.
Despite the growing literature on LAAs and mitigation
techniques a majority of the previous studies have focused
solely on transmission systems and frequency stability except
for which investigated LAAs in distribution systems.
Compared to our work provides analytical insights into the
impact of LAAs and identifies the most vulnerable nodes of
the system which guides the design of the mitigation strategy.
Furthermore our work considers defending against a strategic
and knowledgeable attacker while accounting for errors in the
attack localization and for attacker resource constraints.
In summary there exists a lack of comprehensive investigations into the effects of LAAs on distribution networks
and potential mitigation methods. To address this gap we
first present closedform expressions to determine the voltage
profile of distribution systems under LAA considering ZIP
load models and calculate the minimum required devices to
be manipulated for a successful LAA. This analysis provides
insights into the effect of the attacks spatial location and the
number of devices required to cause the system to violate
voltage safety standards.
Based on these insights we next develop a mitigation
algorithm relying on network reconfiguration that dynamically
alters the spatial location of the attack within the distribution
network topology . A major advantage is that our
proposed mitigation method does not require installing any
additional devices in the system and leverages the existing
flexibility of the network reconfiguration. Additionally to take
into account a strategic attacker and to defend against the
worstcase attacks we model the attackerdefender interaction
using a Stackelberg game framework with the attacker as the
leader and the defender as the follower. The sequential game is
relevant for reactive defenses in which mitigation is triggered
only when an attack is detected note that cyberattacks are
rare events and developing a robust reactive defense strategy
provides the operator significant cost savings. Moreover
drawing from the insights presented in we account
for noisy system measurements and uncertainties associated
with attack detectionlocalization into the utility functions.
Reconfiguration of distribution systems is a wellstudied
research topic. In the reconfiguration is modeled as
a mixed integer secondorder conic programming MISOCP
optimization problem to minimize power losses and increase
system reliability. In this paper reconfiguration is applied in
the context of mitigating cyber attacks by combining it with a
gametheoretic formulation. Since the payoff computation in
the game requires solving the optimization problem repeatedly
we reformulate the network reconfiguration problem as a
mixed integer linear programming MILP optimization by
using the linearized distribution flow LinDistFlow and ZP
approximations. The resulting formulation provides significant
computational speed up in calculating the games payoffs.
Additionally we integrate the ZIP load model into the
constraints to take the voltage dependency of load demand into
account. The proposed mitigation also bears similarities with
the movingtarget defense approach that has received considerable attention in the past few years . However the
proposed strategy applies the reconfiguration reactively rather
than the proactive approach used in these past works and is
specifically tailored to mitigate the effects of LAAs. Fig.
briefly illustrates the proposed gametheoretic interaction of
attacker and defender in this paper.
This paper significantly extends our preliminary work in
terms of using the insights derived from attack impact analysis
to design a mitigation strategy to counter LAAs. The key
contributions are summarized as follows
Deriving closedform expressions to obtain the bus voltages of the distribution system in the presence of voltagedependent loads. Based on these results analysing the effects
of LAAs launched at different locations in the distribution network and obtaining the minimum number of IoTcontrollable
devices required for a detrimental attack.
Proposing a mitigation strategy to counter the effects of
LAA based on the distribution network reconfiguration.
Fig. Summary of the proposed attackerdefender interaction.
Formulating a sequential game approach to the model
strategic attackers while considering the potential uncertainties
associated with the localization of LAAs.
Validating the proposed framework by extensive simulations using the IEEE bus and bus distribution systems.
II. PRELIMINARIES
In this section we introduce the system load and power flow
models for the distribution systems considered in this research.
A. Distribution System Model
The distribution system under study is represented by
the connected directed graph G N L where N
. . . N denotes the set of buses and L denotes the set of
branches. This graph has a radial structure hence it is a tree.
Except for bus which is the root each bus is referred to as
the child of its parent bus which is the adjacent bus closer
to bus by one branch. Thus the set of branches is defined
as L i
i i
i N where i represents the parent
of bus i. In this configuration bus represents the generator
bus. We denote by Dk the set of buses which forms the unique
path connecting bus to bus k excluding bus and including
bus k. The depth of each bus represents the distance in terms
of the number of branches between that bus and the root bus.
B. Load Model
This subsection introduces the load models which are implemented in the rest of the paper.
ZIP Load Model The power demand under the ZIP load
model is given in as follows
s
zip
i
vi p
i
ppvipv
i
jq
i
qqviqv
i
where k k k where k p q. The ZIP load
model captures the voltage dependency of realworld loads.
ZP Approximation Based on the ZIP model is a
function of both vi and v
i
. This causes the optimization tasks
involving the power flow in the presence of ZIP loads to
become nonconvex and complex. To tackle this problem
have provided an approximate model for ZIP loads given by
s
zp
i
vi p
i
p
p
v
i
jq
i
q
q
v
i
where
p p
p
q q
q
p p
p
and
q q
q
while s
zp
i
vi p
zp
i
vi jqzp
i
vi. The new
coefficients in the ZP model are obtained by the binomial
approximation method. The ZP approximation is valid as long
as the voltage is close enough to the nominal value i.e. while
vi vnom . the ZP approximation is valid .
C. Power Flow Equations
Branch Flow Model The branch flow model encapsulates the complete AC power flow with the equations
describing the system state as follows
X
kik
sik sii zii Iii
si
where vi vi zii Iii sii vi
I
ii. Note that
superscript
denotes the conjugate of a complex number.
Linearized Distribution Flow Linearized distribution
flow LinDistFlow simplifies the branch flow model described
in by neglecting branch power losses and is widely
adopted in several distribution grid studies. The power flow
equations under this model are given by
X
kikL
pik pji p
L
i
X
kikL
qik qji q
L
i
v
i v
j rijpji xij qji.
In this formulation and are active and reactive power
balances in each bus and is the equation for finding
subsequent voltage profile.
D. Reconfiguration of Distribution System
Network reconfiguration involves modifying the distribution
networks topology by adjusting the openclosed state of its
switches. In this subsection a set of MILP optimization constraints is introduced primarily to determine the configuration
that maintains the nodal voltages closest to their nominal values. We modify the formulation in to accommodate the ZP
approximation of the loads to capture its voltage dependency.
We also change the power flow model to the LinDistFlow. As
we show in Section V these approximations result
in a MILP the network reconfiguration problem and provide
significant computational speedups compared to the MISOCP.
Connectivity Constraints First we present the connectivity constraints that determine the connection between the
nodes while keeping the overall graph radial
bij
bif f N f
bij bji i j L Ls
bij bji yij i j Ls
yij
X
jijL
bji . i N N f
Equation implies that the substation buses can not have a
parent bus forces the lines without a switch to be always
connected while lets the lines with switches to be either
open or close and forces nonsubstation buses to have
exactly one parent bus.
Power Flow Constraints Below we present the optimization problems power flow constraints which are taken from
the DistFlow and ZP models
pij M bij
qij M bij
X
jijL
pij p
f
i
i N f
X
jijL
qij q
f
i
i N f
X
jN
pji pij p
ZP
i
i N N f
X
jN
qji qij q
ZP
i
i N N f
p
ZP
j p
l
j
v
j
a
p
ZP
j q
l
j
v
j
b
vi
v
i vi
v
ij M bij
v
ij v
i rijpij xij qij
v
i
X
jN
v
ji. i N N f
Equations represent the power flow constraints
is ZIP load constraint and are voltage
constraints. The reason behind using the auxiliary variable of v
is to make the optimization following disciplined convex programming DCP rules of Python. Additionally we consider
all of the normally connected lines to have switches thus they
could be switched off if necessary.
E. LoadAltering Attack Model
Some major manufacturers of highwattage IoTcontrollable
devices have acknowledged the presence of security vulnerabilities in their products . The concept presented in the
LAAs involves attackers leveraging these vulnerabilities to
transform a group of such devices into bots and toggle them
on and off. This coordinated action is intended to disrupt
the stability of the system. Based on this to implement the
LinDistFlow while there is such an attack in the system
and change into
P
jN pji pij p
L
i
i N N f
i N a
P
jN pji pij p
L
i p
a
i N N f
i N a
and
P
jN qji qij q
L
j
i N N f
i N a
P
jN qji qij q
L
j q
a
i N N f
i N a
.
In the following based on the provided models and context
we first analyze the impact of LAA on distribution systems.
Then we propose a gametheoretic mitigation scheme for
these attacks.
III. EFFECTS OF LAA ON DISTRIBUTION SYSTEMS
In this section we analyze the impact of LAA on distribution networks. Our objective is to derive closedform
expressions for the voltage profile of the network with voltagedependant loads under LAA and for the minimum number
of compromised load devices required to cause nodal voltage
safety violations. It is worth noting that system voltages under
LAA can also be computed by solving the power flow equation
through an iterative approach such as the backwardforward sweep BFS technique. However unlike closedform
equations which we derive in this section the application of
iterative techniques does not yield analytical insights into the
impact of LAAs on the distribution network. Furthermore the
closedform expressions obtained in this section are essential
in the design of the defense strategies to mitigate LAAs.
A. Closedform Approximation of Nodal Voltages
To derive the closedform expressions for the system voltages under LAAs we make two approximations i employing
LinDistFlow formulations and ii utilizing the ZP model.
Without LAA First we model the distribution system
without LAA and analyze the power flow equations in
. Integrating into results in
vk
s
v
X
iDk
riip
zp
ii xiiq
zp
ii
where p
zp
ii pii
p
pv
i
q
zp
ii qii
q
qv
i
.
Next we perform a variable change uk v
k
which results
in a set of linear equations which can be written in matrix
form as follows
UN NN
U
N
where U is the vector of squares of voltages and nn
is the matrix with entries
i
X
mDi
rmmp
mm
p xmmq
mm
q
ik
Pi
c rccp
k
p xccq
k
q
if i Dk
ik otherwise
where i k N. We rewrite the system of linear
equations of as
INN
NN UN
N
in which
N k
NN ik for i
... N and k N .
With LAA Here we analyze the voltage profile of the
network under LAA. For this we integrate the introduced LAA
in Section IIE into .
v
a
k
q
v
k pA
a
rka qA
a xka
in which k
P
iDk
riip
zp
i
i xiiq
zp
ii
rka P
iDaDk
rii and xka
P
iDaDk
xii. This
change results in a new set of coefficient matrices. To calculate
the attacked systems square of voltages vector U
A we solve
U
A
N A
NN
U
A
N
.
To obtain the coefficients matrices and are dragged
into LinDistFlow as the ZP model is imposed on them. The
final results are
A
ik ik
a
ik for i and k
A
NN A
N
A
NN in which for i
and k and
a
i
X
cDiDa
p
A
a
p
rcc q
A
a
qxcc
a
ik
riip
a
p xiiq
a
q
if i Da
A
ik otherwise.
B. Analytical Insights into the Attack Impact
Next we use the closedform expressions obtained above
to derive analytical insights into the attack impact. First note
that when there is an attack in a leaf bus the last bus of each
branch rka and xka have the highest possible values. As a
result the voltage drop resulted from p
a
rka q
axka
in is higher and obtained voltages shrink. In conclusion
attacks on the leaf buses yield the most detrimental effects.
While this result is somewhat expected for distribution
systems we further obtain the minimum number of attacked
devices which leads to voltage safety violations. We call such
a threat the critical attack. For this we consider the voltage
of the leaf bus as a known variable vth. The new unknown
variable is p
a
and based on attacked device type we can find
q
a via q
a
qd
pd
p
a
. So the new set of coefficients for obtaining
voltages of buses except for the leaf one and the active power
of the critical attack is forming
d
d d
d
in which
d
i
i v
thia if i a
i v
thia if i a
d
ik
P
cDi
rcc
p
qd
pd
xcc
q
if k a i Da
d
ik if k a i Da
ik otherwise.
Hence we can solve the linear system of equations given by
I
a
d
X d
in which X is a vector with the same dimension as U where
all elements of it are the same as U except for one in which X
contains p
a
instead of ua since we already know ua v
th.
Additionally I
a
is the identity matrix except for the element
a a which equals . This gives the bus voltages as well a
the required p
a
. Then we can find the number of required
devices in bus n using cn p
a
pd
.
The closedform expressions provided in this section are
not only used for identifying the worst effect of the LAAs
but also in the following sections they inspire our mitigation
method to find the optimal defensive action. Additionally the
introduced critical attack has been implemented in finding
the optimal action of the strategic attacker later in this paper.
IV. MITIGATING LAA VIA RECONFIGURATION
In this section we introduce a novel technique to mitigate LAAs by reconfiguring the distribution system topology.
We exploit a sequential gametheoretic interaction in which
following the LAA launched by the attacker the defender
reconfigures the network to react optimally to the threat.
A. Mitigation Design and Intuition
The intuition behind the proposed defense technique of
reconfiguring the distribution network lies in the analytical
insights derived in Section III. Based on our analysis recall
that LAAs targeting the leaf buses of the distribution network
lead to the greatest attack impact in terms of the deviation
of the voltage from the nominal values. For instance as
represented in Fig. in the base configuration of the IEEE
bus system an LAA targeting node the leaf bus of the
longest branch is the most impactful attack. Note that in Fig.
dashed lines present the normally open lines disconnected
lines which could be connected in case of reconfiguration.
In this context reconfiguring the distribution networks
topology e.g. closing the link between Bus and Bus
in bus grid changes the position of the leaf buses thus
alleviating the attack impact. It is worth noting that the
proposed mitigation leverages the preexisting capabilities
of the distribution network e.g. devices enabling network
reconfiguration are primarily installed to reduce power losses
andor voltage deviations and hence does not require new
infrastructure.
Furthermore in the proposed scheme the system will be
reconfigured only when an attack is detected thus avoiding
unnecessary reconfigurations to mitigate attacks note that
cyberattacks are somewhat rare events. The attack detection
module can be based on existing modeldriven or datadriven
approaches for detecting LAAs. The reader can refer to past
works including and in this area for more details.
B. Stackelberg Game for Attack Mitigation
We model the strategic interaction of the attacker and the
defender via a noncooperative game. Under this formulation
the attacker first chooses a bus to launch an LAA and
subsequently the defender exploits the flexibility of the system
to reconfigure it and mitigate the attack. This successive
interaction is modeled by a Stackelberg game. The sequential
game models this reactive mitigation framework.
A Stackelberg game with two players consists of a leader
and a follower. The leader always commits their action first
to maximize their own objective function by anticipating the
followers strategic reaction. Then given the action of the
leader the follower picks their best i.e. optimal response
to maximize their own objective function. Since LAAs are
rare incidents in the network we propose a reactive mitigation
method instead of a preventive one. In this approach the
defender waits for the attacker and responds optimally.
Our Stackelberg game can be formally defined as H
A DSA SDFA FD in which A and D are the
players attacker and defender SA SD denote the set of
available actions of players and FA FD denote their reward
or payoff functions.
The attackers set of actions SA is launching an LAA
in any of the buses one attack on one bus at a time by
assumption that causes voltage constraint violation. Thus the
set of actions of the attacker is SA N L in which N L N
denotes the set of load buses.
The set of defense actions is the set of all possible system
reconfigurations discussed in Section IID such that SD
. . . NB denotes the set of indices of all possible
reconfiguration matrices B bij ijN
NN
whose entries meet the constraints in equations and
NB represents the number of such matrices. Henceforth we
denote by Bd bij dijN the adjacency matrix of the
network as a result of a specific defense d SD.
The attacker wants to maximize the voltage deviation as a
result we define
FAd a X
nN
v
nom v
nd a
as the attackers objective function. The reason behind using
the squares of the voltages in the above equation is that it
greatly simplifies the computation of the Stackelberg equilibrium involving a MILP as will be discussed later. Moreover
increasing the distance between the squares of voltages is
equivalent to increasing the distance between the voltages
given that vn vnom hence leading to a consistent
reward for the attacker.
The defenders objective is to reconfigure the system to
minimize the square of voltage deviation above. Achieving
this goal with minimal changes can decrease the maintenance
requirements for switches and reduce the likelihood of switching failures. To take this factor into account we add a penalty
term and the resulting reward function of the defender is
F
perf
D
X
nN
v
nom v
nd a
pend
where the first part of the F
perf
D represents the sum of the
square of voltage deviations in all buses and
pend X
N
i
X
N
j
b
pre
ij bij d
is the penalty term for enforcing a system reconfiguration with
the minimum switches possible.
The above reward F
perf
D is relevant when the defender is
capable of perfect attack localization. However due to noisy
measurements the defender might not be able to do this.
Instead we assume that the defender is only able to locate
a neighbourhood of the attack i.e. a connected cluster of
buses which contains the bus under attack. We further assume
that the defender has a favourite candidate bus under attack
denoted by natt but does not discard the attack possibilities of
other buses in the located cluster. To model this we define a
discrete probability vector . . . N with entries
if natt
Anatt
if Anatt
otherwise.
Above represents the likelihood that the defender assigns
to bus
P
being under attack such that and
N . Additionally . denotes the likelihood
of the defenders favourite candidate. The subset Anatt is the
set of buses directly adjacent to natt along with the buses
directly adjacent to those all these buses are considered as
the other potential candidates by the defender. Their likelihood
is the remaining probability split equally between the
Anatt other candidate buses.
For instance if in the base configuration of the system of
Fig. the defender chooses as favourite candidate bus natt
with a likelihood of then A and
the probabilities are .
. and N A.
Taking this uncertainty of detecting precisely the attack
location into account at the defenders end results in the
following reward
FDd a
X
N
X
nN
v
nom v
nd a
pend
which represents the expected reward over this uncertainty.
Obtaining the optimal attack and defense requires computing the Stackelberg equilibrium. As discussed earlier in
this paper the attacker commits the attack first and then the
defender reacts.
Definition . The best response of the defender to an action
a SA is defined as
ra arg max
dSD
FDd a.
Definition . A profile of actions a
d
SA SD is a
Stackelberg equilibrium iff
FAra
a
FAra a a SA
d
ra
.
Intuitively the attacking action at the Stackelberg equilibrium is the one which maximizes the attackers reward when
the defender responds with the best reaction. Furthermore the
defenders best reaction to a
is its Stackelberg equilibrium
action.
A Stackelberg equilibrium is ensured to exist if the defenders optimal response exists for every action of the leader.
Assuming that normally open points exist in the distribution
system this ensures that at least one system reconfiguration is
possible and that the discrete feasible set in is nonvoid
leading to the existence of the solution.
Algorithm describes the method of finding the players
actions at the Stackelberg equilibrium.
First we need to find the best reaction ra for each attack
a SA which requires running a reconfiguration optimization
Algorithm Computing the Stackelberg equilibrium.
Data A DSA SDFA FD
... N
Result ra
a
Compute ra a SA by solving
ra arg maxdSD FDd a s.t.
and as a MILP optimization problem via
SCIPY and CVXPY
Compute FAra a a SA
Choose a
arg maxaSA FAra a
Compute ra
arg maxdSD FDd a
s.t.
and by solving the defined MILP
optimization problem
for each of the attacks see equation . Note that the
constraints of this optimization are linear in the square of the
voltages hence justifying our choice of the distance between
the squares of voltages in the reward function leading to
a linear program instead of a quadratic one obtained by
a simple variable change ui v
i
. The resulting MILP
optimizations are carried out in Python and the optimization
modeling language is CVXPY.
The solver used for these optimizations is SCIPY. Then
find the reward of the attacker for ra a a SA. Finally
the action corresponding to the maximum of FAra a is
selected as the attacking strategy in our Stackelberg formulation. Additionally ra corresponding to the attacking action
at the Stackelberg equilibrium a
is the optimum defensive
strategy at the Stackelberg equilibrium d
ra
.
C. ResourceConstrained Attacker
Drawing from our discussions in Section IIIB it is evident
that each bus has a distinct critical attack leading to voltage
constraint violation. Given the attackers tendency for launching such a critical attack their potential action will not only
occur across different buses but also vary in magnitude.
To accommodate this feature we define a new game H
A DSA SDF
A FD in which the attacker launches
the critical attack which we call them the resourceconstrained attacker. Note that the critical attack in each bus
is the minimum number of required devices to be manipulated
in that bus to cause voltage constraint violation. Indeed cn
for each attack is unique and varies with the attack location.
In this modified game F
A comprises two components the
total nodal voltage deviation and the attack magnitude. The
attacker seeks to maximize the former while minimizing the
latter. However these two terms cannot be simply summed
due to their disparate physical characteristics. Therefore we
propose the following reward
F
Ad a F
norm
A d a cnorma
in which and
F
norm
A d a FAd a
P
iSA
FAd i
c
norma ca
P
iN L ci
.
TABLE I Number of compromised devices required to cause
voltage safety violations in the IEEE bus network during different
hours of the day.
Time Att. location
bus
Least load Peak
load
Air
Conditioner
Resistive
heater
Note that the parameter trades off between the two
components of the objective function. If the attacker
only cares about maximizing the harm caused in the voltage
profile similar to the case using and if the
attacker only cares about minimizing the attacked devices.
The rest of the components of H
are the same as H. The
process of computing the Stackelberg equilibrium is similar to
H and we only need to plug in the attackers new objective
function F
d a.
V. RESULTS AND DISCUSSIONS
Our simulations are conducted using the IEEE bus and
bus systems. The ZIP load coefficients of all the buses are
set to the residential loadtype F introduced in .
A. Critical Attack
Here we conduct spatial analyses to determine the most
effective location for launching LAAs. To quantify this we use
the load profile obtained from for as the base
load without LAAs which contains the hourly electric power
demand in New York US. To mimic this load profile in the
IEEE bus test system we project the ratio of load changes
at different hours of the day onto the nominal load of the
test network. Table I presents the number of devices required
for the critical attack in three of the leaf buses of the bus
test case during the different hours of the day. These numbers
are computed via the equations in Section IIIB. We can see
that the attack on the bus with the highest depth requires
fewer devices to be manipulated. This conclusion confirms
our insights from Section IIIA. Furthermore it also shows
the dependency on the type of load air conditioner resistive
load etc. and the associated ZIP load coefficients.
Since the results in Table I are obtained by the approximation discussed in Section III we evaluate how effective these
attacks are when considering the full AC power flow model.
To evaluate the extent of the errors between the voltages
computed using the analytical results and the full AC power
flow model we compare the obtained voltage profile with
the results of BFS. Fig. shows the voltage profile calculated
by the two methods during the peak load demand and the
corresponding critical attack. We remark that the actual errors
of computing the nodal voltages with our approximations
never exceed thus proving the validity of the analytical
results.
B. LAA Mitigation
Next we examine the proposed LAA mitigation method.
The base load profiles without LAAs of bus and bus
test cases are and of their nominal load profiles
Fig. Voltage profile of the attacked on Bus bus test case
with the proposed closedform equations and the accurate model.
TABLE II Normally open lines.
bus
bus
obtained by MATPOWER. Except for the resourceconstrained
attacker results we consider LAA attacks of magnitude p
a
i
q
a
i . p.u. in which i is the index of the attacked bus.
This magnitude of the attack is significant enough to cause
voltage safety violations in all the load buses except those
adjacent to the root and hence needs to be mitigated. Table
II represents the candidate lines which can be openedclosed
during the system reconfiguration in the two bus systems.
Three scenarios are considered next i accurate attack
localization ii errors in attack localization and iii errors in
attack localization and resourceconstrained attacker.
i Accurate Attack Localization Here we set natt
and N L N a
. The results for the bus grid
are presented in Tables III and IV. In all figures red buses
represent the attacked bus attackers action and green lines
are the ones with altered state been opened or closed as the
defenders reaction. Based on the results the attack is launched
on Bus . Fig. shows the voltage profiles of the grid under
attack before and after reconfiguration. The result shows that
the proposed mitigation method is able to return the voltage
profile within the constraints hence mitigate the effects of
LAA. Although the LAA on Bus causes the greatest impact
in terms of voltage deviations a strategic attacker that can
anticipate the defenders action chooses to launch the attack
on Bus instead to maximize their payoff.
Fig. and Table III illustrate the players actions at the
Stackelberg equilibrium for the bus grid in which attack is
launched on Bus and lines and changed
their state to reconfigure the network.
ii Errors in Attack Localization In this scenario we
consider . of certainty about the location of the attack then the remaining . probability of the attack location
is split equally between the other buses in the neighbourhood
see Section IV. The result for the bus grid is presented in
Tables III and IV. In this scenario the defense action should
keep the voltage within the desired constraints assuming an
LAA in any of the candidate buses. This uncertainty causes a
system reconfiguration that necessitating more switching. Fig.
represents the voltage profile of the bus grid after the
reconfiguration if any of the three suspect buses are attacked.
We can notice that the attack will be successfully mitigated
and the voltage constraint will not be violated in any case.
Note that the players actions for the bus system are the
same as in scenario i.
Fig. Voltage profile of the attacked bus grid before and after
mitigation reconfiguration.
Fig. IEEE bus system configuration after the attack and
reconfiguration.
iii ResourceConstrained Attacker Finally we present
the resourceconstrained attacker introduced in Section IVC
with . in their objective function. The uncertainty
level in the attack localization is the same as in the previous
scenario. The result for the bus system is presented in
Tables III and IV. We observe that in this scenario the attacker
chooses to attack Bus as the impact of load alteration on
this bus will be the greatest as the number of compromised
devices is taken explicitly into account. The new configuration
selected by the defender and illustrated in Table IV tackles the
voltage constraint violation by this attacker.
Again the bus case maintains the same Stackelberg
equilibrium actions as the past two scenarios.
Table V represents the total voltage deviations
P
iN vnom vi
in the grid for different scenarios.
Note that in scenario ii the voltage deviation of the bus
grid drops but the defender needs to commit more switching
which is not desirable. Furthermore a portion of the reduced
voltage deviation in scenario iii should be attributed to the
smaller attack launched by the resourceconstrained attacker.
C. Significance of GameTheoretic Approach
We also consider a nonstrategic attacker that does not
anticipate the defenders actions. In this case first the attacker
launches the attack which maximizes the total voltage deviation regardless of potential defensive reactions. Subsequently
the defender solves the optimal reconfiguration problem to
mitigate the attack. The results of this approach for the
bus test case are presented in Table VI. We remark that both
scenarios i and ii result in the same output. Compared to
Table V for the strategic attacker the defender always benefits.
Indeed in scenario i the total voltage deviation is dropped
from . p.u. to . p.u. while in ii the number of
switching is reduced from to . Note that mitigation with
Fig. Voltage profiles of bus grid attacked at any of suspect
buses after mitigation.
TABLE III Attacked bus and closedopen lines at the Stackelberg
equilibrium of each scenario.
Scenario Attacked bus Closed lines Opened lines
bus bus bus bus bus bus
i
ii
iii
less switching is preferred even with slightly higher voltage
deviation.
Note that the strategic attacker in the bus grid picks the
deepest bus Bus as the victim which is similar to the
nonstrategic attackers choice.
D. Approximation Efficiency
Finally to show the efficiency of the approximations introduced in Section IID we compare the computation time for
the proposed MILP model proposed in IID with that of the
similar MISOCP model. Exhibited in Table VII for the
bus test case a single MISOCP optimization takes . s
while one MILP optimization requires . s which means
a reduction of over in calculation time while for the
bigger system this reduction is even more significant .
It is important to note that as we aim to find the Stackelberg
equilibrium solving the optimization must be iterated for each
potential attack to uncover all potential rewards for players.
Thus the proposed approximations provide significant computation time reduction to compute the Stackelberg equilibrium.
VI. CONCLUSIONS
In this paper we propose a set of closedform expressions
for the power flow of distribution systems to determine bus
voltages in the presence of voltagedependent loads with or
without LAA which we exploit to investigate the impact
of LAAs. Then we introduce a sequential gametheoretic
approach to mitigate LAAs in distribution systems in a reactive
manner by network reconfiguration with minimum possible
switching. Furthermore we take into account the uncertainties in the attack localization by introducing a probability
distribution over the potentially attacked nodes. Finally by
introducing a hybrid objective function we considered the
resourceconstrained attacker.
REFERENCES
S. Maleki S. Pan E. V. Belmega C. Konstantinou and S. Lakshminarayana The impact of load altering attacks on distribution systems
with ZIP loads arXiv preprint arXiv. .
TABLE IV Stackelberg equilibrium configurations of the bus grid Red buses are the location of the attack.
Scenario i Accurate attack localization Scenario ii Error in localization the attack Scenario iii Resourceconstrained attacker
TABLE V Total voltage deviations and switching required under
each scenario.
Scenario Number of switching Total voltage deviation p.u.
bus bus bus bus
i . .
ii . .
iii . .
TABLE VI Nonstrategic attackers preferred actions obligated
switching numbers and total voltage deviations.
Scenario Attacked bus Total voltage
deviation p.u.
Number of
switching
i .
ii .
S. Amini F. Pasqualetti and H. MohsenianRad Dynamic load altering
attacks against power system stability Attack models and protection
schemes IEEE Transactions on Smart Grid vol. no. pp.
.
A.H. MohsenianRad and A. LeonGarcia Distributed internetbased
load altering attacks against smart power grids IEEE Transactions on
Smart Grid vol. no. pp. .
S. Soltan P. Mittal and H. V. Poor BlackIoT IoT botnet of high
wattage devices can disrupt the power grid in th USENIX Security
Symposium pp. .
S. Lakshminarayana S. Adhikari and C. Maple Analysis of IoTbased
load altering attacks against power grids using the theory of secondorder
dynamical systems IEEE Transactions on Smart Grid vol. no.
pp. .
A. Dabrowski J. Ullrich and E. R. Weippl Grid shock Coordinated
loadchanging attacks on power grids The nonsmart power grid is
vulnerable to cyber attacks as well in Proceedings of the rd Annual
Computer Security Applications Conference pp. .
Z. Liu and L. Wang A robust strategy for leveraging soft open points
to mitigate load altering attacks IEEE Transactions on Smart Grid
vol. no. pp. .
S. Acharya Y. Dvorkin and R. Karri Public plugin electric vehicles
grid data Is a new cyberattack vector viable? IEEE Transactions on
Smart Grid vol. no. pp. .
B. Huang A. A. Cardenas and R. Baldick Not everything is dark and
gloomy Power grid protections against IoT demand attacks in USENIX
Security Symposium pp. .
J. Ospina X. Liu C. Konstantinou and Y. Dvorkin On the feasibility of loadchanging attacks in power systems during the COVID
pandemic IEEE Access vol. pp. .
S. Lakshminarayana J. Ospina and C. Konstantinou Loadaltering
attacks against power grids under COVID lowinertia conditions
IEEE Open Access Journal of Power and Energy vol. pp.
.
Y. Dvorkin and S. Garg IoTenabled distributed cyberattacks on
transmission and distribution grids in North American Power
Symposium NAPS. IEEE pp. .
M. P. Goodridge S. Lakshminarayana and A. Zocca Uncovering loadaltering attacks against N secure power grids A rareevent sampling
approach IEEE Transactions on Power Systems arXiv.
.
S. Soltan P. Mittal and H. V. Poor Protecting the grid against MAD
attacks IEEE Transactions on Network Science and Engineering vol.
no. pp. .
A. Di Giorgio A. Giuseppi F. Liberali A. Ornatelli A. Rabezzano and
L. R. Celsi On the optimization of energy storage system placement
for protecting power transmission grids against dynamic load altering
attacks in Proc. Mediterranean Conference on Control and Automation.
IEEE pp. .
TABLE VII Optimization times in seconds.
Op. type bus grid bus grid
MISOCP . .
MILP . .
Z. Chu S. Lakshminarayana B. Chaudhuri and F. Teng Mitigating
loadaltering attacks against power grids using cyberresilient economic
dispatch IEEE Transactions on Smart Grid vol. no. pp.
.
M. A. Sayed R. Atallah C. Assi and M. Debbabi Electric vehicle attack impact on power grid operation International Journal of Electrical
Power Energy Systems vol. p. .
M. A. Sayed M. Ghafouri R. Atallah M. Debbabi and C. Assi
Protecting the future grid An electric vehicle robust mitigation scheme
against load altering attacks on power grids Applied Energy vol.
p. .
Y. Guo L. Wang Z. Liu and Y. Shen Reinforcementlearningbased
dynamic defense strategy of multistage game against dynamic load
altering attack International Journal of Electrical Power Energy
Systems vol. p. .
H. Haghighat and B. Zeng Distribution system reconfiguration under
uncertain load and renewable generation IEEE Transactions on Power
Systems vol. no. pp. .
S. F. Santos M. Gough D. Z. Fitiwi J. Pogeira M. Shafiekhah and
J. P. Catalao Dynamic distribution system reconfiguration considering
distributed renewable energy sources and energy storage systems IEEE
Systems Journal vol. no. pp. .
H. Jahangir S. Lakshminarayana C. Maple and G. Epiphaniou A
deeplearningbased solution for securing the power grid against load
altering threats by iotenabled devices IEEE Internet of Things Journal
vol. no. pp. .
Q. Li J. Zhang J. Zhao J. Ye W. Song and F. Li Adaptive hierarchical
cyber attack detection and localization in active distribution systems
IEEE Transactions on Smart Grid vol. no. pp. .
R. S. Rao S. V. L. Narasimham M. R. Raju and A. S. Rao Optimal
network reconfiguration of largescale distribution system using harmony search algorithm IEEE Transactions on Power Systems vol.
no. pp. .
T. Van Cutsem and C. Vournas Voltage stability of electric power
systems. Springer Science Business Media .
S. Lakshminarayana E. V. Belmega and H. V. Poor Movingtarget
defense against cyberphysical attacks in power grids via game theory
IEEE Transactions on Smart Grid vol. no. pp. .
C. Liu J. Wu C. Long and D. Kundur Reactance perturbation for
detecting and identifying FDI attacks in power system state estimation
IEEE Journal of Selected Topics in Signal Processing vol. no.
pp. .
F. U. Nazir B. C. Pal and R. A. Jabr Approximate load models for
conic OPF solvers IEEE Transactions on Power Systems vol. no.
pp. .
M. Farivar and S. H. Low Branch flow model Relaxations and
convexificationpart I IEEE Transactions on Power Systems vol.
no. pp. .
M. Baran and F. F. Wu Optimal sizing of capacitors placed on a radial
distribution system IEEE Transactions on Power Delivery vol. no.
pp. .
J. A. Taylor and F. S. Hover Convex models of distribution system
reconfiguration IEEE Transactions on Power Systems vol. no.
pp. .
A. Bokhari A. Alkan R. Dogan M. DiazAguil o F. De Leon
D. Czarkowski Z. Zabar L. Birenbaum A. Noel and R. E. Uosef
Experimental determination of the ZIP coefficients for modern residential commercial and industrial loads IEEE Transactions on Power
Delivery vol. no. pp. .
New York Independent System Operator NYISO load data
httpswww.nyiso.comloaddata accessed . |
|
write a youtube script | Im making a tabletop bowling lane with an automatic pin Setter last time we built a lock and place latch mechanism for the bumpers my actual Lane was just these big D printed blocks I installed alignment pins to minimize the lip between each block somehow I always start with the worst ideas first I dont know why a better idea would just be to use some wood so I went to Lowes and picked up a T long plank I then D printed these risers that attached to it I made these Corner blocks to align the wood to the plas I then flipped the whole thing over and drilled some pilot holes I then installed some threaded inserts into the wood Im not really a wood guy I dont know what the hell Im doing with those inserts installed I can now easily attach my plastic parts to the plank and it makes it easy to take apart frequently after the risers were attached it was time to build out two sets of the bumper mechanisms this part was freaking tedious but we got it done then I attached the bumpers to the Lane I also added some gutters I think Im going to add some mesh or something on the inside of the bumpers because the pins were getting stuck under them Im really excited to build out the automated pin Setter next especially because manually resetting the pins for each shot right now is not fun follow to see the rest of my build |
|
write a youtube script | "LaRa Efficient LargeBaseline Radiance Fields\nAnpei Chen Haofei Xu Stefano Esposito\n\nSiyu Tang\n (...TRUNCATED) |
|
write a youtube script | "arXiv.v cs.CC Jul \nDetermination Problems for Orbit Closures and Matrix Groups\nRida Ait El Manss(...TRUNCATED) |
|
write a youtube script | "Topological Generalization Bounds for DiscreteTime\nStochastic Optimization Algorithms\nRayna Andre(...TRUNCATED) |
End of preview. Expand
in Dataset Viewer.
README.md exists but content is empty.
Use the Edit dataset card button to edit it.
- Downloads last month
- 43