学术报告 | Secure Self-supervised Learning

浙江大学网络空间安全学院

学术报告

  摘 要  

Self-supervised learning (SSL) is an emerging machine learning (ML) paradigm, which relies on unlabeled datasets to pre-train powerful encoders that can then be treated as feature extractors for various downstream tasks. Despite being powerful, SSL is also vulnerable to various security and privacy attacks. In this talk, I will summarize some of our projects covering both attacks and defenses, with particular focus on membership/attribute inference attacks(CCS 2021), more effective model stealing attacks (Preprint), and copyright protection (CCS 2022). I will wrap up with a discussion of open directions on this topic.

  报告人简介  

Xinlei He is a second-year Ph.D. student at CISPA Helmholtz Center for Information Security advised by Prof. Yang Zhang. His research focuses on trustworthy machine learning, misinformation, hateful speech, and memes. He has published over 10 papers in top-tier conferences/journals. He served as the TPC of ESORICS 2021 (poster session) and ESORICS 2022. He is the recipient of The Norton Labs Graduate Fellowship 2022. More details are at http://www.xinlei.info/.