Diffusion Models For Adversarial Machine Learning And Social Good

The main three research objectives of this project revolve around producing state of the art generative AI techniques (using diffusion models) for adversarial machine learning , with an emphasis on social good. Our first objective is to develop novel adversarial attacks on facial recognition systems using diffusion models.

These attacks will result in stealthy adversarial props that can be physically realisable as well as being robust to different circumstances such as different lighting and poses. Secondly, to develop novel adversarial attacks for images that prevent the malicious use of AI-powered editing tools.

Current countermeasures have flaws such as not being robust to transformations, assuming an unrealistic white box scenario, requiring a large amount of memory to generate and are not successful every time.

Developing novel countermeasures would ensure that the countermeasures keep up to date with the ever changing editing tools as well as having the desirable properties discussed. Thirdly, to develop novel adversarial attacks for videos that prevent the malicious use of AI-powered editing tools.

While there exists several adversarial attacks for different social good scenarios as discussed earlier, there is significantly less research on preventing the malicious use of AI-powered editing tools for videos. Developing novel countermeasures for videos therefore would be significant and socially impactful.