Safety of GenAI through the Lens of Security and Cryptography

In this deliberately provocative two-part talk from the recent workshop on Theoretical Aspects of Trustworthy AI, Somesh Jha (University of Wisconsin) makes a case for applying a security and cryptography mindset to evaluating the trustworthiness of machine learning systems, particularly in adversarial and privacy-sensitive contexts. He contrasts the rigorous threat modeling and attack-defend cycles of security research with the often insufficient evaluation standards in mainstream ML, especially regarding robustness, privacy, and watermarking. Watch Part I here and Part II here.