Verifiable Privacy and other Constraints in Mobile Device GenAI

Abstract

Federated learning (FL) is a field of ML focused on learning from decentralized data (e.g., across a population of mobile phones). FL at Google has undergone a transformation over the last several years, in order to (a) enable privacy protections which are verifiable by external parties and (b) support LLM-powered data science workflows. In this talk I'll present an overview of our new FL paradigm which leverages trusted execution environments ('TEEs'). I'll share what TEE-based analytics and learning workflows are now enabled, as well as research threads motivated by these new capabilities. Finally, I'll share work by my teammates and I outside of (and complementary to) privacy, on efficient on-device ML.

Speaker bio

Sean is a Staff Research Scientist at Google Research. His research interests are in maximizing the utility and efficiency of AI models in highly constrained scenarios and environments. At present, that manifests as work in mobile device AI on topics like federated learning and LLM personalization, under constraints like verifiable user privacy, limited on-device computation, and limited server communication. In a previous life, this manifest as work on optimizing satellites for various space applications. He has a PhD from Stanford University in Aeronautics & Astronautics.