[2019-12-31] Mr. Greg Yang, Microsoft Research AI, "Building Provably Robust Image Classifiers via Randomized Smoothing and Adversarial Training"

Poster:Post date:2019-12-23
Title: Building Provably Robust Image Classifiers via Randomized Smoothing and Adversarial Training
Date: 2019-12-31 14:20pm-15:30pm
Location: R210, CSIE
Speaker: Mr. Greg Yang, Microsoft Research AI
Hosted by: Prof. Yen-Huan Li


It is now well-known that deep neural networks suffer from adversarial attacks: An adversary can add a small change in an input image imperceptible to humans, but it results in dramatic change in a neural network’s classification of the image. In the last few years, a significant amount of work has gone into designing defenses, but all except a notably few cases are broken by stronger or more clever attacks later on. In response, many works have considered *provable defenses* which produce certificates that the defense cannot be broken. We report on our latest work that obtains the state-of-the-art robust accuracy on CIFAR-10 and Imagenet against L2-norm-bounded adversaries. Our technique involves smoothing a neural network with Gaussian noise and then training this *smoothed* classifier adversarially. Our paper is available at arxiv.org/abs/1906.04584 and our code is available at github.com/Hadisalman/smoothing-adversarial.
Greg Yang is a researcher at Microsoft Research AI in Redmond, Washington. He joined MSR after he obtained Bachelor's in Mathematics and Master's degrees in Computer Science from Harvard University, respectively advised by ST Yau and Alexander Rush. He won the Hoopes prize at Harvard for best undergraduate thesis as well as Honorable Mention for the AMS-MAA-SIAM Morgan Prize, the highest honor in the world for an undergraduate in mathematics, for his work on homological theory of functions. He has given invited talks at many institutions, with the most recent being the International Congress of Chinese Mathematicians 2019, the most premier mathematics conference in China. 

Last modification time:2019-12-23 AM 11:19

cron web_use_log