Akademik Veri Yönetim Sistemi
Journal of Supercomputing, cilt.81, sa.18, 2025 (SCI-Expanded, Scopus)
Recommender systems personalize product recommendations based on users’ past purchasing and browsing interactions. Advanced recommender systems in a multi-criteria context enhance the process by providing more personalized and accurate recommendations. However, they are vulnerable to shilling attacks, where fake profiles are introduced to manipulate recommendations. A specific type of shilling attack, the Power User Attack (PUA), involves cloning influential users to distort recommendations. While the susceptibility of traditional recommender systems to such attacks has been widely studied, the resilience of these advanced recommender systems against PUAs remains unexplored. This study designs PUA variants for the multi-criteria context and evaluates the robustness of state-of-the-art techniques, including deep learning-based approaches, under multiple power user identification strategies and attack scenarios. The experimental grid is combinatorial and computationally intensive. To make this evaluation tractable on a large scale, we use a parallelized evaluation pipeline for evaluated recommendation techniques and utilize GPU-accelerated training for deep models. We analyze prediction shifts caused by PUAs, propose two novel robustness metrics, and assess attack effectiveness via ANOVA and Kruskal–Wallis tests on real-world datasets. Our findings provide actionable guidance for selecting and hardening multi-criteria-oriented advanced recommenders in adversarial environments and encourage the development of future robustness-aware designs.