Trust and power-aware group shilling attacks on multi-criteria top-N recommendation systems

Multi-Criteria Collaborative Filtering (MCCF) systems improve recommendation quality by considering users’ preferences across multiple criteria. However, the rich multidimensional structure of MCCF systems also makes them vulnerable to sophisticated shilling attacks. In particular, coordinated group shilling attacks can generate behaviorally informed fake profiles capable of significantly manipulating recommendation results. Although various attack generation methods have been proposed in the literature, existing approaches generally rely on random filler item selection and fail to effectively exploit the behavioral characteristics of influential users in multi-criteria environments. To address this limitation, this study proposes a novel trust- and power-aware group shilling attack model called Group Shilling Attack Power Attack (GSAGenPA) . Unlike traditional attack models, the proposed approach identifies powerful users based on user influence metrics and generates attack profiles by imitating their preference behaviors. In addition, new trust-oriented target item selection strategies are introduced for both push and nuke attack scenarios. The proposed framework is evaluated on multi-criteria recommendation datasets under different attack configurations. Experimental results demonstrate that GSAGenPA provides a strategically informed attack-generation mechanism that consistently improves recommendation manipulation performance over the random filler-selection baseline ( \(GSAGenₗ Ran\) ). However, the proposed power-aware variants do not generally surpass the average-based filler-selection baseline ( \(GSAGenₗ Avg\) ) in terms of absolute attack success, highlighting a trade-off between power-aware profile construction and maximum manipulation effectiveness.