An Unseen Threat in Cross-Domain Recommendation: User Re-Identification Attacks and Practical Defenses
Abstract
Cross-domain recommendation (CDR) boosts personalization by transferring knowledge across domains, but its privacy implications remain largely overlooked. In this paper, we uncover a critical threat in CDR: cross-domain user re-identification can be performed in realistic black-box settings using only the top-K recommended items returned by deployed systems. We propose a list-supervised shadow-model attack that treats the target model's top-K lists as implicit feedback and mimics its ranking behavior through a pairwise ranking objective. To amplify identity signals that persist across domains---even with disjoint item sets---the attacker learns disentangled user interests, separating domain-specific preferences from domain-invariant representations through contrastive constraints. The resulting representations support an effective classifier for determining whether two cross-domain accounts belong to the same individual. Extensive experiments across multiple CDR settings show that re-identification leakage is pervasive in modern models. Finally, we present practical defense mechanisms that mitigate linkability while preserving recommendation utility, highlighting the need for privacy-aware CDR design. The codes are available here.