[AI Security and Privacy Team Seminar] Talk by Prof. Kobbi Nissim

RIKEN AIP's AI Security and Privacy Team hosts a seminar by Prof. Kobbi Nissim of Georgetown University, one of the originators of differential privacy and a Gödel Prize laureate for his work in cryptographic theory. The talk, titled "Protecting the Undeleted in Machine Unlearning," examines how the legal Right to be Forgotten under regulations such as the EU GDPR and the California Consumer Privacy Act has driven the field of machine unlearning, which aims to remove specific records from a trained model as if they had never been included. Prof. Nissim will show that even a task computable with strong privacy guarantees can be attacked: any perfect retraining mechanism can let an adversary controlling only a few data points reconstruct nearly the entire dataset simply by issuing deletion requests. He will then present a cryptographically motivated security definition that protects undeleted data points from leakage caused by deleting others, while still supporting essential functionalities such as bulletin boards, summations, and statistical learning. The seminar is held at the Ookayama campus of the Institute of Science Tokyo and streamed online via Zoom. The work is joint with Aloni Cohen, Refael Kohen, and Uri Stemmer.