RPDN: An Effective Rating Pollution Attacks Detection Framework for Recommendation Systems

Yongrui Gu, Jianke Yu, Yongye Li, Peiting Xie, Xiaoyang Wang, Chen Chen

Abstract: With the rapid growth of e-commerce, safeguarding against malicious attacks on their platforms is essential for reliable operation. Among the various existing attack methods, Rating Pollution Attacks (RPAs) seriously affect the recommendation system and reduce the platform’s trading volume. Attackers pose as legitimate users to manipulate or distort the ratings of high-exposure items, creating a substantial challenge for detection. Such attacks can severely mislead platform recommendation systems, resulting in considerable negative consequences for user trust and platform integrity. Existing research focuses on fine-grained attack behavior detection and lacks algorithms specifically designed to detect such attacks. Motivated by the limitation of the existing models, we propose a novel semi-supervised attack detection network named Rating Pollution Detection Network (RPDN). Three well-designed modules, including topology analysis, information propagation and attacker community detection combined with contrastive learning strategies, ensure RPDN ’s excellent information exploitation and attack detection capabilities. Extensive experiments for RPA detection on five popular signed bipartite graph datasets demonstrate RPDN ’s superior performance, achieving an average increase of 13.53% in AUC and 21.96% in Binary F1-Score compared to the second-best method.

External IDs:dblp:conf/pakdd/GuYLXWC25