Membership Inference Attacks Against Cross-Domain Recommender Systems

Cross-domain recommender systems can effectively address the data sparsity and cold-start problems in traditional recommender systems. However, such systems may inadvertently leak user membership privacy due to their reliance on shared latent representations and behavioral patterns across domains. In this paper, we propose a membership inference attack (MIA) framework tailored for cross-domain recommender systems, termed CDRMIA, to quantify privacy risks. Unlike prior MIA methods designed for single-domain settings, our attack infers a user's membership status in both the source and target domains using only target-domain recommendations and partial behavioral data. We design cross-domain features based on embedding alignment, behavioral consistency, and prediction confidence, and employ a multi-task learning model to jointly infer membership states. Extensive experiments on two real-world datasets (Amazon and Douban) and three representative cross-domain recommendation models (EMCDR, CMF, CLFM) demonstrate the effectiveness of CDRMIA. Our findings reveal significant privacy leakage risks in cross-domain recommender systems, highlighting the urgent need for robust privacy-preserving mechanisms.