Answers are grounded in this paper and cite their sources.
Real-time personalized recommendation technology is widely used in daily life. Although federated recommender systems enhance privacy protection by keeping sensitive user data locally, their decentralized training process remains vulnerable to poisoning attacks by malicious clients. Existing poisoning attacks mainly exploit adversarial samples, malicious gradients, or user-item relationships, but their effectiveness may be limited when the proportion of Byzantine clients is small or the candidate-item pool is large. To address these challenges, this paper proposes UdaAttack, a poisoning attack method for federated recommender systems based on user-distribution similarity. The proposed method constructs user distribution models for items and uses distributional similarity to guide item replacement, thereby degrading recommendation performance. In addition, a real-time task scheduling strategy is introduced to improve the efficiency of server-side gradient aggregation. Experiments on the MovieLens-1M dataset demonstrate that UdaAttack can effectively degrade recommendation performance with a relatively low proportion of Byzantine clients.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: