Grant: Research Tools and Instruments

AI models show a worrying susceptibility to adversarial attacks, in which an attacker applies imperceptible changes to the input to arbitrarily influence a target model. In particular, such attacks can jailbreak aligned foundation models. Given the increasing impact of AI models on our society, it is urgent to develop safer models with robust predictions. We propose a novel technique combining denoising diffusion and differential privacy to design AI models that are provably robust against adversarial attacks. Our approach will improve on existing Randomized Smoothing defences, and enable new capabilities such as joint robustness under multiple threat models and adaptivity to the difficulty of the input. However, our recent progress towards this direction showed that our lack of adequate compute resources is hindering the project's development speed, and the quality of our final models. The equipment we request aims to fill three urgent needs to increase research productivity in creating adversarially robust AI models: (1) training HQP in the development of parallelized AI model training; (2) supporting the fast iterative development of AI software that can scale to multiple GPUs; (3) training larger models to showcase state-of-the-art results and increase the impact of our research.