Analysis of Methods, Metrics, and Directions in Shilling Attack Detection using Machine Learning Algorithms

Shilling attacks undermine the integrity of recommender systems, compromising both user trust and system performance. This paper provides a comprehensive review of shilling attack detection methods, focusing on detection metrics and attack dimensions. We analyze 62 studies published between 2000 and 2024, categorizing them based on detection methodologies, attack models, and the algorithms used. Key contributions of the review include a detailed classification of attack detection metrics into generic, model-specific, intra-profile, and residual-based categories, as well as a synthesis of the most commonly applied detection techniques, including supervised, unsupervised, and hybrid models. Our findings reveal that while supervised learning methods, such as Support Vector Machines (SVM) and k-Nearest Neighbors (KNN), dominate the field, deep learning-based approaches and ensemble methods are gaining traction due to their high accuracy in detecting complex attack patterns. Additionally, we identify gaps in current research, particularly the need for more scalable and adaptable detection mechanisms as recommender systems evolve. The scalability of various shilling attack detection methods has been analyzed, and the correlation between detection metrics, attack models, and algorithms has been investigated. This study provides deeper insights for selecting optimal detection techniques in future research. The paper concludes with future directions for enhancing the effectiveness and robustness of shilling attack detection.