Speaker:Tianyu Pang

Affiliation:Sea AI Lab, Singapore

Date and Time:March 27, 2024 (Wednesday), 15:00-16:00

Location:Room 3, 1st Floor, Block A, Jade Science and Education Building

Abstract:

This report will discuss how to mitigate the robust-accuracy trade-off in traditional classification tasks and enhance adversarial robustness through data generation. Additionally, it will explore different implementations of adversarial attacks (jailbreaks) on large multimodal models, including test-time backdoor attacks and infectious jailbreaks. The report will adopt a combined online and offline approach.

Speaker's Bio:

Tianyu Pang is a Senior Research Scientist at Sea AI Lab, Singapore. He graduated from the Special Pilot Program in Mathematics and Physics at Tsinghua University in 2017 and obtained his Ph.D. in Computer Science from Tsinghua University in 2022 under the supervision of Professor Jun Zhu. His primary research interests include trustworthy machine learning and generative models. He has published over 30 papers at ICML, NeurIPS, ICLR, CVPR, ICCV, and ECCV, with a total citation count exceeding 7500. His research has been featured in Forbes and WIRED, and he has consistently ranked in the top three in international AI security competitions. He has received scholarships and awards from Microsoft, Baidu, NVIDIA, WAIC, Zhong Shimu Foundation, CAAI, and the National Scholarship.