Date & Time:
February 18, 2021 1:00 pm – 2:00 pm
Location:
Live Stream
02/18/2021 01:00 PM 02/18/2021 02:00 PM America/Chicago Hongyang Zhang (TTIC) – New Advances in (Adversarially) Robust and Secure Machine Learning Live Stream

New Advances in (Adversarially) Robust and Secure Machine Learning

Watch via live stream.

Deep learning models are often vulnerable to adversarial examples. In this talk, we will focus on robustness and security of machine learning against adversarial examples. There are two types of defenses against such attacks: 1) empirical and 2) certified adversarial robustness.

In the first part of the talk, we will see the foundation of our winning system, TRADES, in the NeurIPS’18 Adversarial Vision Challenge in which we won 1st place out of 400 teams and 3,000 submissions. Our study is motivated by an intrinsic trade-off between robustness and accuracy: we provide a differentiable and tight surrogate loss for the trade-off using the theory of classification-calibrated loss. TRADES has record-breaking performance in various standard benchmarks and challenges, including the adversarial benchmark RobustBench, the NLP benchmark GLUE, the Unrestricted Adversarial Examples Challenge hosted by Google, and has motivated many new attacking methods powered by our TRADES benchmark.
 
In the second part of the talk, to equip empirical robustness with certification, we study certified adversarial robustness by random smoothing. On one hand, we show that random smoothing on the TRADES-trained classifier achieves SOTA certified robustness when the perturbation radius is small. On the other hand, when the perturbation is large, i.e., independent of inverse of input dimension, we show that random smoothing is provably unable to certify L_infty robustness for arbitrary random noise distribution. The intuition behind our theory reveals an intrinsic difficulty of achieving certified robustness by “random noise based methods”, and inspires new directions as potential future work.

Host: Ben Zhao

Hongyang Zhang

Postdoc Fellow, Toyota Technological Institute at Chicago

Hongyang Zhang is a Postdoc fellow at Toyota Technological Institute at Chicago, hosted by Avrim Blum and Greg Shakhnarovich. He obtained his Ph.D. from CMU Machine Learning Department in 2019, advised by Maria-Florina Balcan and David P. Woodruff. His research interests lie in the intersection between theory and practice of machine learning, robustness and AI security. His methods won the championship or ranked top in various competitions such as the NeurIPS’18 Adversarial Vision Challenge (all three tracks), the Unrestricted Adversarial Examples Challenge hosted by Google, and the NeurIPS’20 Challenge on Predicting Generalization of Deep Learning. He also authored a book in 2017.

Related News & Events

UChicago CS News

New Assistant Professor Rana Hanocka Combines AI, 3D, and Computer Graphics

Feb 09, 2022
In the News

Quanta Magazine Features Prof. Bill Fefferman’s Work on Quantum Algorithms

Jan 20, 2022
UChicago CS News

UChicago CS Prof. Ben Zhao Named ACM Fellow

Jan 19, 2022
UChicago CS News

CS 4th Year Sophie Veys Receives CRA Undergraduate Research Award

Jan 14, 2022
UChicago CS News

In-Fridge Controller Could Scale Up Quantum Computers, Award-Winning UChicago Research Finds

Jan 10, 2022
UChicago CS News

UChicago Workshop Highlights Internet Frontiers and Opportunities

Dec 02, 2021
UChicago CS News

Prof. Rebecca Willett Named IEEE Fellow

Nov 29, 2021
UChicago CS News

Aaron Elmore Promoted to Associate Professor at UChicago Computer Science

Nov 24, 2021
UChicago CS News

ScaleStuds Project Receives $5 Million to Build Foundations for Massive Computation

Nov 19, 2021
In the News

Asst. Prof. Blase Ur Discusses the “Metaverse” on Chicago Tonight

Nov 05, 2021
UChicago CS News

Using AI and Data Science to Reliably Detect Internet Censorship in Real Time

Nov 02, 2021
UChicago CS News

EPiQC Research Receives Best Paper Award at IEEE Quantum Week

Oct 22, 2021
arrow-down-largearrow-left-largearrow-right-large-greyarrow-right-large-yellowarrow-right-largearrow-right-smallbutton-arrowclosedocumentfacebookfacet-arrow-down-whitefacet-arrow-downPage 1CheckedCheckedicon-apple-t5backgroundLayer 1icon-google-t5icon-office365-t5icon-outlook-t5backgroundLayer 1icon-outlookcom-t5backgroundLayer 1icon-yahoo-t5backgroundLayer 1internal-yellowinternalintranetlinkedinlinkoutpauseplaypresentationsearch-bluesearchshareslider-arrow-nextslider-arrow-prevtwittervideoyoutube