
Trust and Distrust in ML: Privacy, Verification and Robustness (Part 2)
Shafi Goldwasser, director of the Simons Institute for the Theory of Computing and professor at UC Berkeley, MIT, and the Weizmann Institute, continues her Emmy Noether Lecture on trust in machine learning systems. This second part builds on foundational cryptographic concepts to examine how privacy guarantees, verification protocols, and robustness properties can be formally established for ML models. Goldwasser addresses how adversaries might manipulate or extract information from models, and what cryptographic tools, such as zero-knowledge proofs and differential privacy, can offer provable guarantees against such attacks. The talk, delivered at the Institute for Advanced Study's Wolfensohn Hall, connects her decades of work in complexity theory and cryptography to urgent open problems in trustworthy AI, including how to verify that a model was trained correctly and how to detect tampering or backdoors. The lecture assumes some familiarity with cryptographic and machine learning concepts from the series' first part.