Visiting Researcher Talk: Prof Yan Wang | 27 Jan 2026

Talk Title
Membership Inference Attacks in Recommender System

Speaker

Prof Yan Wang

About the Speaker

Dr. WANG Yan is currently a Full Professor in the School of Computing, Macquarie University, Australia. He is also the Research Director of Macquarie University Centre for Frontier AI Research (FAIR). He received his PhD from Harbin Institute of Technology (HIT), P. R. China. Prior to joining Macquarie University in 2003, he was a Postdoctoral Fellow/Research Fellow in the Department of Computer Science, School of Computing, National University of Singapore (NUS). He has published a number of research papers in international conferences including AAAI, AAMAS, CVPR, ICDE, IJCAI, KDD, NeurIPS, SIGIR, WWW, and journals including CSUR, TIST, TKDD, TKDE, TOIS, TSC and TWEB. In addition, the proposed solution on dual-target cross-domain recommendation systems has been adopted by Alipay system. His research interests cover recommender systems, fake news detection/mitigation, data analytics, trust management and social computing.

Description

Recommender systems have been successfully applied in many applications, such as movie recommendation in Netflix. However, not much attention has been paid to the membership inference attacks (MIAs) where an attacker aims to infer whether a user’s data has been used to train a target recommender system. In this talk, we will introduce two pieces of our work investigating MIAs on recommender systems.

In the first work, we propose a novel shadow-free MIA that directly leverages a user's recommendations for membership inference. Without shadow training, the proposed MIA can be conducted efficiently and effectively under a practice scenario where the attacker is given only black-box access to the target recommender system.

In the second work, we study MIAs against hybrid-based recommender systems where the same algorithm utilises user-item historical interactions and attributes of users and items to produce personalised recommendations for both existing users and new users. Specifically, we propose a novel metric-based MIA to leverage the characteristics of personalisation to obtain reference recommendations for a target user. Then, a relative membership metric is proposed to exploit a target user’s historical interactions, target recommendations, and reference recommendations to infer the membership of the target user’s data.