StegaAttnGAN: Lightweight Cross-Attention Steganography

StegaAttnGAN is an end-to-end generative adversarial framework for image steganography — hiding discrete token sequences (e.g., text) inside natural images.
It integrates:

• A Transformer-based message encoder
• A U-Net style generator with multi-scale cross-attention fusion
• An 8×8 spatial memory Transformer decoder
• A lightweight discriminator
• A stochastic noise layer for robustness against compression and perturbations

The project was developed with the assistance of Generative AI and is released as open source for research and educational purposes.

So Called Paper: https://github.com/Andreas-UI/StegaAttnGAN/blob/main/StegaAttnGAN.pdf

---

Features

• Hide short messages in small images (e.g., CIFAR-10, 32×32).
• Multi-scale cross-attention for adaptive message embedding.
• Curriculum training on message length and objective weighting.
• Robust decoding under common distortions (Gaussian noise, blur, down/up sampling, JPEG-like quantization).
• Open source and extensible for future research.

---

Architecture

---

Results

Message Length	PSNR (dB) ↑	Image Loss (MSE) ↓	Text Loss (CE) ↓
3	26.76	0.0085	1.96

---

Installation

Clone the repository and install dependencies:

git clone https://github.com/Andreas-UI/StegaAttnGAN.git
cd StegaAttnGAN
pip install -r requirements.txt

Requirements include:

• Python 3.10+
• PyTorch (CUDA recommended)
• torchvision
• tqdm
• numpy
• matplotlib

Usage

Training

To start training on CIFAR-10 with curriculum learning:

python training_example.py

This will:

• Train the model across message lengths (1 → max_msg_len).
• Alternate between text-focused and image-focused phases.
• Save checkpoints and log training progress.

Inference

To run inference with a trained model (trained weights):

python inference_example.py

This will:

• Load a pretrained checkpoint.
• Encode a random message into a cover image.
• Decode the message from the stego image.
• Print accuracy and visualize cover/stego/residual images.

Disclaimer

This project was developed with the help of Generative AI. The author provides the code and documentation as-is and assumes no responsibility for errors, inaccuracies, or misuse. Use at your own risk.

Contributing

Contributions are welcome! Please open an issue or submit a pull request for:

• Bug fixes
• New features
• Improved documentation
• Additional datasets or evaluation pipelines