Hanshen Xiao, When is Automatic Privacy Proof Possible for Black-Box Processing?

Can we automatically and provably quantify and control the information leakage from a black-box processing? From a statistical inference standpoint, in this talk, I will start from a unified framework to summarize existing privacy definitions based on input-independent indistinguishability and unravel the fundamental challenges in crafting privacy proof for general data processing. Yet, the landscape shifts when we gain access to the (still possibly black-box) secret generation. By carefully leveraging its entropy, we unlock the black-box analysis. This breakthrough enables us to automatically "learn" the underlying inference hardness for an adversary to recover arbitrarily-selected sensitive features fully through end-to-end simulations without any algorithmic restrictions. Meanwhile, a set of new information-theoretical tools will be introduced to efficiently minimize additional noise perturbation assisted with sharpened adversarially adaptive composition. I will also unveil the win-win situation between the privacy and stability for simultaneous algorithm improvements. Concrete applications will be given in diverse domains, including privacy-preserving machine learning on image classification and large language models, side-channel leakage mitigation and formalizing long-standing heuristic data obfuscations. About the speaker: Hanshen Xiao is an Assistant Professor in the Department of Computer Science. He received his Ph.D. degree in computer science from MIT and B.S. degree in Mathematics from Tsinghua University. Before joining Purdue, he was a research scientist at NVIDIA Research. His research focuses on provable trustworthy machine learning and computation, with a particular focus on automated black-box privatization, differential trust with applications on backdoor defense and memorization mitigation, and trustworthiness evaluation.