Experimental and Theoretical Study for the Popular Shilling Attacks Detection Methods in Collaborative Recommender System

The stability and reliability of filtration and recommender systems are crucial for continuous operation. The presence of fake profiles, known as “shilling attacks,” can undermine the reliability of these systems. Therefore, it is important to detect and classify these attacks. Numerous techniques for detecting shilling attacks have been proposed, including supervised, semi-supervised, unsupervised, Deep Learning, and hyper deep learning methods. These techniques utilize well-known shilling attack models to target collaborative recommender systems. While previous research has focused on evaluating shilling attack strategies from a global perspective, considering factors such as attack size and attacker’s knowledge, there is a lack of comparative studies on the various existing and commonly used attack detection methods. This paper aims to fill this gap by providing a comprehensive survey of shilling attack models, detection attributes, and detection algorithms. Furthermore, we explore the traits of injected profiles that are exploited by detection algorithms, which has not been thoroughly investigated in prior works. We also conduct experimental studies on popular attack detection methods. Our experimental results reveal that hybrid deep learning algorithms exhibit the highest performance in shilling detection, followed by supervised learning algorithms and semi-supervised learning algorithms. In contrast, the unsupervised technique performs poorly. The deep learning-based Shilling Attack Detection demonstrates accuracy and quality in accurately identifying a variety of mixed attacks. This study provides valuable insights into shilling attack models, detection attributes, and detection algorithms. Our findings highlight the superior performance of hybrid deep learning algorithms in shilling detection, as well as the limitations of unsupervised techniques. Deep learning-based Shilling Attack Detection showcases its effectiveness and accuracy in identifying various types of attacks.