
Trust and Distrust in ML: Privacy, Verification and Robustness
Shafi Goldwasser, director of the Simons Institute for the Theory of Computing and professor at UC Berkeley and MIT, delivers this Emmy Noether Lecture at the Institute for Advanced Study on the gap between machine learning's growing power and its trustworthiness. She lays out three technical pillars for building confidence in learned models: privacy guarantees that keep training data from leaking back out of a model, verification methods that let a weaker party check a claimed computation or prediction without redoing the work, and robustness techniques that keep models reliable under adversarial inputs or distribution shift. Drawing on cryptographic tools she helped develop over her career, including zero-knowledge proofs, Goldwasser connects decades-old theory to present-day concerns about deploying ML in high-stakes settings. The talk is aimed at researchers in theoretical computer science and cryptography but stays accessible to anyone following the debate over whether black-box models can be made accountable.