Crypto Seminar Talk with Andrea Rizzini
Info about event
Time
Location
Nygaard-295, Finlandsgade 21-23, 8200 Aarhus N
Title: Verifiable Integrity Claims for Privacy-Preserving Federated Learning
Abstract: Federated learning is usually presented as a privacy-friendly way to train models on data that never leaves its owner. Most of the literature, however, assumes that the parties involved actually follow the protocol, that's a reasonable assumption when the clients are hospitals or banks, much less so in open, general-purpose deployments. If we assume clients and the aggregator as untrusted, a range of attacks becomes possible: poisoned updates, tampering with the global model, silently dropping contributions by the aggregator, free riding by the clients. In this talk I argue that verifiability- evidence that a computation was carried out as declared, obtained without redoing it - is the missing ingredient, and that it can be layered on top of existing privacy protections rather than traded against them. I will lay out a framework that describes a verifiable FL round in terms of a transcript and a set of integrity claims, covering both the client side and the aggregator side, and show how different deployments arise depending on who plays the role of verifier. I will then use this framework to systematize thirteen recent works, based on either zero-knowledge proofs or trusted execution environments, and discuss what the comparison reveals.
About the Speaker: Andrea Rizzini is a PhD Student at the Polytechnic University of Milan, Italy. This week, he is visiting the Cryptography and Cybersecurity Section.