BEGIN:VCALENDAR
PRODID:-//eluceo/ical//2.0/EN
VERSION:2.0
CALSCALE:GREGORIAN
BEGIN:VEVENT
UID:www.tcs.tifr.res.in/event/1397
DTSTAMP:20240228T090542Z
SUMMARY:Towards Robust and Reliable Machine Learning: Adversaries and Funda
 mental Limits
DESCRIPTION:Speaker: Arjun Bhagoji (University of Chicago)\n\nAbstract: \nW
 hile ML-based AI systems are increasingly deployed in safety-critical sett
 ings\, they continue to remain unreliable under adverse conditions that vi
 olate underlying statistical assumptions. In my work\, I aim to (i) unders
 tand the conditions under which a lack of reliability can occur and (ii) r
 eason rigorously about the limits of robustness\, during both training and
  test phases.In the first part of the talk\, I demonstrate the existence o
 f strong but stealthy training-time attacks on federated learning\, a rece
 nt paradigm in distributed learning. I show how a small number of compromi
 sed agents can modify model parameters via optimized updates to ensure des
 ired data is misclassified by the global model\, while bypassing custom de
 tection methods. Experimentally\, this model poisoning attack leads to a l
 ack of reliable prediction on standard datasets.Test-time attacks via adve
 rsarial examples\, i.e. imperceptible perturbations to test inputs\, have 
 sparked an attack-defense arms race. In the second part of the talk\, I st
 ep away from this arms race to provide model-agnostic fundamental limits o
 n the loss under adversarial input perturbations. The robust loss is shown
  to be lower bounded by the optimal transport cost between class-wise dist
 ributions using an appropriate adversarial point-wise cost\, the latter of
  which can be efficiently computed via a linear program for empirical dist
 ributions of interest.To conclude\, I will discuss my ongoing efforts and 
 future vision towards building continuously reliable and accessible ML sys
 tems by accounting for novel attack vectors and new ML paradigms such as g
 enerative AI\, as well as developing algorithmic tools to improve performa
 nce in data-scarce regimes.\nShort Bio:\nArjun Bhagoji is a Research Scien
 tist in the Department of Computer Science at the University of Chicago. H
 e obtained his Ph.D. in Electrical and Computer Engineering from Princeton
  University\, where he was advised by Prateek Mittal. Before that\, he rec
 eived his Dual Degree (B.Tech+M.Tech) in Electrical Engineering at IIT Mad
 ras\, where he was advised by Andrew Thangaraj and Pradeep Sarvepalli. Arj
 un's research has been recognized with a Spotlight at the NeurIPS 2023 con
 ference\, the Siemens FutureMakers Fellowship in Machine Learning (2018-20
 19) and the 2018 SEAS Award for Excellence at Princeton University. He was
  a 2021 UChicago Rising Star in Data Science\, a finalist for the 2020 Bed
 e Liu Best Dissertation Award in Princeton's ECE Department and a finalist
  for the 2017 Bell Labs Prize.\n
URL:https://www.tcs.tifr.res.in/web/events/1397
DTSTART;TZID=Asia/Kolkata:20240229T113000
DTEND;TZID=Asia/Kolkata:20240229T123000
LOCATION:A-201 (STCS Seminar Room)
END:VEVENT
END:VCALENDAR