---
license: mit
language:
- ja
pipeline_tag: image-to-text
---
# Steganography Neural Network Model Card

## Model Overview
**Task:** Image Steganography (Message Embedding and Extraction)  
**Architecture Type:** Encoder-Decoder Neural Network  
**Primary Use Case:** Embedding and recovering hidden messages in images  

## Technical Specifications
- **Parameters:** 980,548
- **Model Size:** 3.74 MB
- **Precision:** torch.float32
- **FLOPs:** 1,954,176
- **Input Resolution:** 512 × 512 pixels
- **Framework:** PyTorch

## Architecture Details

### Encoder Network
- **Input:** 4 channels (RGB + message), 512×512px
- **Output:** 3 channels (RGB stego image), 512×512px
- **Key Components:**
  - Initial Conv (4→64 channels)
  - Backbone with SE blocks and dilated convolutions
  - Residual connections
  - Final weighted combination (0.9 × original + 0.1 × encoded)

### Decoder Network
- **Input:** 3 channels (stego image), 512×512px
- **Output:** 1 channel (recovered message), 512×512px
- **Key Components:**
  - Feature extraction (3→64→128 channels)
  - SE blocks and residual connections
  - Message extraction pathway

## Training Details
- **Hardware:** GTX 1080 GPU
- **Epochs:** 600
- **Optimizer:** AdamW (lr=0.001, weight_decay=0.01)
- **Scheduler:** Cosine Annealing (min_lr=1e-6)
- **Loss Functions:**
  - Image Loss: 0.95×MSE + 0.05×(1-SSIM)
  - Message Loss: MSE
  - Combined with dynamic alpha weighting

## Key Features
- Group Normalization for batch-size independence
- SiLU activation functions throughout
- Squeeze-and-Excitation blocks for channel attention
- Dilated convolutions in encoder
- Skip connections for detail preservation

## Performance Characteristics
- Maintains visual image quality while embedding messages
- Optimized for both image fidelity and message recovery
- Lightweight architecture (<1M parameters)

## Limitations and Biases
- Fixed input resolution of 512×512 pixels


## Technical Requirements
- PyTorch environment
- GPU recommended for optimal performance
- Standard deep learning dependencies
- Sufficient memory for 3.74 MB model

## Citation and Contact
- Model source and citation information not provided
- Contact information for maintainers not specified