Omar Montasser: Adversarially Robust Learning: Characterization, Reductions, and Robustness to Unknown Perturbations

Speaker: Omar Montasser , TTIC

Date: Wednesday, March 24, 2021

Time: 12:00 PM to 1:00 PM Note: all times are in the Eastern Time Zone

Public: Yes

Location: Zoom: https://mit.zoom.us/j/99597752648

Event Type: Seminar

Room Description:

Host: David Sontag, CSAIL

Contact: Hunter Lang, hjl@mit.edu

Relevant URL:

Speaker URL: https://ttic.uchicago.edu/~omar/

Speaker Photo:
None

Reminders to: cdml-students@mit.edu, cdml-core@mit.edu, seminars@csail.mit.edu, mitml@mit.edu

Reminder Subject: TALK: Adversarially Robust Learning: Characterization, Reductions, and Robustness to Unknown Perturbations

Abstract:
We study the problem of learning an adversarially robust predictor from i.i.d. training data. This talk will focus on two questions. First, can we learn adversarially robust predictors using a black-box non-robust learning algorithm? We give a reduction algorithm for robustly learning any class H using any non-robust PAC learner for H, with nearly-optimal oracle complexity. Second, can we design learning algorithms with robustness guarantees without knowing the perturbation set a-priori? We examine a model where a learning algorithm is allowed to query an attack oracle, but does not know the perturbation set otherwise. We show that a class H is robustly learnable in this model if and only if H is online learnable.
Based on joint works with Steve Hanneke and Nathan Srebro: https://arxiv.org/abs/2010.12039 and https://arxiv.org/abs/2102.02145.
Speaker bio:
Omar Montasser is a fourth year PhD student at TTI-Chicago advised by Nathan Srebro. His main research interest is the theory of machine learning. Recently, his research focused on understanding and characterizing adversarially robust learning, and designing algorithms with provable robustness guarantees under different settings. His work has been recognized by a best student paper award at COLT (2019).

Research Areas:
AI & Machine Learning

Impact Areas:

This event is not part of a series.

Created by David A. Sontag Email at Tuesday, March 16, 2021 at 5:43 PM.