BEGIN:VCALENDAR VERSION:2.0 PRODID:-//wp-events-plugin.com//7.2.3.1//EN TZID:Asia/Kolkata X-WR-TIMEZONE:Asia/Kolkata BEGIN:VEVENT UID:107@cds.iisc.ac.in DTSTART;TZID=Asia/Kolkata:20250224T110000 DTEND;TZID=Asia/Kolkata:20250224T120000 DTSTAMP:20250217T134418Z URL:https://cds.iisc.ac.in/events/seminar-cds-102-february-24th-1100-towar d-realizing-user-level-differential-privacy-at-scale/ SUMMARY:{Seminar} @ CDS: #102\, February 24th\, 11:00: "Toward realizing us er-level differential privacy at scale" DESCRIPTION:Department of Computational and Data Sciences\nDepartment Semin ar\n\n\n\nSpeaker : Prof. Krishna Pillutla\, Assistant Professor\, IITM\nT itle : "Toward realizing user-level differential privacy at scale"\nDate & amp\; Time : February 24\, 2025 (Monday)\, 11:00 AM\nVenue : # 102\, CDS S eminar Hall\n\n\n\nABSTRACT\n\nThere is a growing realization that in-doma in user data is crucial to unlocking the full potential of AI models. Howe ver\, use of such data comes at the cost of increased risk of leaking info rmation and compromising the privacy of individual users. In this talk\, I 'll present some building blocks required to realize user-level differenti al privacy (DP) to protect the privacy of all the (possibly related) examp les contributed by any individual user. We focus on some fundamental quest ions: Why do we even need user-level DP? How do we attain it and audit it? \n\nWe start by showing that an adversary can reliably infer whether a use r's data was used in training. Crucially\, this is possible using only a f ew fresh samples of the user's data that were not used for training. Heuri stic mitigation strategies have limited success\, motivating the need for the strong guarantees provided user-level DP.\n\nIn the next part of the t alk\, we look at private learning algorithms\, focusing on a class of nois y stochastic gradient algorithms that inject temporally correlated noise. While these algorithms enjoy the state-of-the-art utility\, they suffer fr om a quadratic runtime complexity. We improve the runtime complexity to ne arly linear at no cost in the privacy-utility tradeoff both in theory (wit h near optimal error bounds) and in practice (with significant empirical i mprovements).\n\nThe final step of a practical (user-level) DP implementat ion is an empirical audit to verify the correctness of the claimed DP guar antee. We present a randomized auditing procedure that significantly reduc es the number of training runs to give a high probability lower bound on t he privacy leakage. Along the way\, we introduce a benchmark of large-scal e user-stratified datasets to enable investigations into user-level DP and federated learning at the scale of foundation models.\n\nI'll conclude wi th a set of future research directions and concrete applications. This is based on collaborative work with many and will touch upon results from the following publications: EMNLP (2024\; Oral)\, FOCS (2024)\, ICLR (2024)\, NeurIPS (2023)\, NeurIPS D&\;B (2023)\, SaTML (2025)\n\nBIO: Krishna P illutla is an assistant professor at the Wadhwani School of Data Science a nd AI at IIT Madras in India. Previously\, he has been a visiting research er (postdoc) at Google Research in the Federated Learning team. He obtaine d his Ph.D. at the University of Washington where he was advised by Zaid H archaoui and Sham Kakade. Before that\, he received his M.S. from Carnegie Mellon University and B.Tech. from IIT Bombay.\n\nKrishna's research has been recognized by a NeurIPS outstanding paper award (2021)\, a JP Morgan Ph.D. fellowship (2019-20)\, and two American Statistical Association (ASA ) Student Paper Award Honorable Mentions.\n\nHost Faculty: Dr. Danish Prut hi\n\n\n\nALL ARE WELCOME CATEGORIES:Events,Talks END:VEVENT BEGIN:VTIMEZONE TZID:Asia/Kolkata X-LIC-LOCATION:Asia/Kolkata BEGIN:STANDARD DTSTART:20240225T110000 TZOFFSETFROM:+0530 TZOFFSETTO:+0530 TZNAME:IST END:STANDARD END:VTIMEZONE END:VCALENDAR