Date: Wednesday, March 20th, 2024
9:00 am – 10:00 am Pacific Time
12:00 pm – 1:00 pm Eastern Time
Location: Weekly Seminar, Zoom
Title: Posthoc Privacy guarantees for neural network queries
Abstract:
Cloud based inference in machine learning is an emerging paradigm where users share their data with a service provider. Due to increased concerns over data privacy, recent works have proposed using Adversarial Representation Learning (ARL) to learn an informal privacy-preserving encoding of sensitive user data before it is shared with an untrusted service provider. Traditionally, the privacy of these encodings is evaluated empirically as they lack formal guarantees. Our work helps add formal privacy guarantees to such informally private pipelines that share embeddings by interjecting with a special processing that is done Posthoc-and hence the name Posthoc privacy. In this work, we develop a new framework that provides formal privacy guarantees for an arbitrarily trained neural network by linking its local Lipschitz constant with its local sensitivity. To utilize local sensitivity for guaranteeing privacy, we extend the Propose-Test-Release(PTR) framework to make it tractable for neural network-based queries. We verify the efficacy of our framework experimentally on real-world datasets and elucidate the role of ARL in improving the privacy-utility tradeoff.
Bio:
Praneeth Vepakomma is currently an Assistant Professor at MBZUAI. Prior to that, he obtained his PhD at MIT. His research interests are in trustworthy, responsible machine learning and collaborative data science. He has extensive industrial experience across Meta, Apple, Amazon, Motorola Solutions, Corning and several startups. He won the ADIA Lab Fellowship, Meta PhD research fellowship in Applied Statistics and two SERC Scholarships (for Social and Ethical Responsibilities of Computing) from MIT’s Schwarzman college of computing. His non-profit won the Financial Times Digital Innovation Award. He won a Best Student Paper Award at FL-IJCAI, a Baidu Best Paper Award at NeurIPS-SpicyFL and a Best Paper Runner Up Award at FG-2021. His technical work is inspired by foundations of non-asymptotic statistics, randomized algorithms, federated computation, privacy, data valuation methods and at times just by systems design. He has organized several workshops at ICLR, ICML, IJCAI, CVPR and NeurIPS.