Zoom: https://illinois.zoom.us/j/2432644784?pwd=M0NYZkpIUThXM0I2bCtpcUxYbHJjZz09
Abstract:
Along with the wide deployment of deep learning (DL) systems, their lack of trustworthiness (robustness, fairness, numerical reliability, etc) is raising serious social concerns, especially in safety-critical scenarios such as autonomous driving, aircraft navigation, and facial recognition. Hence, a rigorous and accurate evaluation of the trustworthiness of DL systems is critical before their large-scale deployment.
In this talk, I will introduce my research on certifying critical trustworthiness properties of large-scale DL systems. Inspired by techniques in optimization, cybersecurity, and software engineering, my work computes rigorous worst-case bounds to characterize the degree of trustworthiness for a given DL system and further improve such bounds via strategic training. Specifically, I will introduce two representative frameworks: (1) DSRS is the first framework with theoretically optimal certification tightness. DSRS along with our training method DRT and accompanying open-source tools (VeriGauge and alpha-beta-CROWN) is the state-of-the-art and award-winning solution for achieving DL robustness against constrained perturbations. (2) TSS is the first framework for building and certifying large DL systems with high accuracy against semantic transformations. TSS opens a series of subsequent research on guaranteeing semantic robustness for various downstream DL and AI applications. I will conclude this talk with a roadmap that outlines several core research questions and future directions on trustworthy machine learning.
Bio:
Linyi Li is a Computer Science PhD candidate at UIUC, advised by Prof. Bo Li and co-advised by Prof. Tao Xie. Prior to his PhD, Linyi Li earned his bachelor’s degree in Computer Science from Tsinghua University in 2018. His research lies in the intersection of computer security, machine learning, and software engineering. He focuses on building certifiably trustworthy deep learning systems at scale by proposing state-of-the-art certification and training methods for various trustworthy properties such as robustness, fairness, and numerical reliability. He has published over 20 papers at S&P, CCS, ICML, NeurIPS, ICLR, ICSE, FSE, etc. He is the main developer or key contributor of several widely-known and award-winning deep learning certification tools including alpha-beta-CROWN (winner of VNN-COMP 2022), VeriGauge, CROP, and COPA. Linyi is a recipient of Adversarial Machine Learning Rising Star Award, Rising Star in Data Science Award, the Wing Kai Cheng Fellowship, and a finalist of Qualcomm Innovation Fellowship and Two Sigma PhD Fellowship.
https://wiki.illinois.edu/wiki/display/grad/PILOT+SEMINAR
