Generative Adversarial Masks: Safeguarding Images from AI Models with Stealthy Perturbations

6.8301 Project. By Daniel Prakah-Asante and Aileen Liao

Project Overview

"Generative Adversarial Masks" (SPAM) is a research project focused on protecting images from being analyzed by AI models. The project develops a Stealth Perturbation Adaptive Model (SPAM) that applies image-specific perturbations to protect images from AI analysis while maintaining visual similarity.

Key Features

• Protection through Stealth Perturbations: Masks generated using the SPAM model safeguard images from AI models by altering their representations.
• Flexible Protection Levels: The model allows for adjustable parameters to balance between image quality and protection based on the use case.
• Evaluation on Multiple Datasets: Tested on both Flickr30k and Dogs vs Cats datasets, demonstrating reduced AI classifier performance.

Models

Several variations of the SPAM model were implemented:

• Scaled Epsilons: Adjusts the perturbation strength (ε) during training to control the ratio between image quality and protection.
• Norm Models: Minimize additional norms to improve image quality.
• Secondary Loss Models: Introduce gradient and magnitude losses to better preserve image structure while enhancing protection.
• Alternative Loss Models: Focus on collapsing the image's representation into uniform vectors.

Evaluation

• Image Quality: The Structural Similarity Index Measure (SSIM) and Cosine Similarity metrics are used to evaluate the balance between protection and visual fidelity.
• Model Robustness: Testing on the Dogs vs Cats dataset demonstrated that SPAM-protected images reduced AI classifier accuracy from 99% to 48%.

Results

• The RGB-Gradient Model showed the best performance in preserving image quality, achieving an average SSIM of 0.99.
• More aggressive protection models significantly degraded the performance of AI classifiers without sacrificing image quality.

Conclusions

The SPAM model effectively protects images from AI analysis, balancing protection with image quality. The RGB-Gradient method proved most effective for applications where image quality is a priority. Future work will explore scaling the model to protect against more complex AI architectures and utilizing diffusion techniques for robust image protection.

Future Work

• Expand the model to protect against advanced architectures like CLIP.
• Investigate the use of diffusion models for creating more robust protective masks.

Acknowledgments

Special thanks to the MIT staff for their guidance.

License

This project is licensed under the MIT License.

Citation

If you use or reference this project, please cite it as follows:

@misc{spam_protection_2024,
  title={Generative Adversarial Masks: Safeguarding Images from AI Models with Stealthy Perturbations},
  author={Prakah-Asante, Daniel and Liao, Aileen},
  year={2024},
  publisher={GitHub},
  journal={GitHub repository},
  howpublished={\url{https://github.com/username/generative-adversarial-masks}}
}