Fusing hypergraph spectral features for shilling attack detection. (December 2021)

• Record Type: Journal Article
• Authors: Li, Hao; Gao, Min; Zhou, Fengtao; Wang, Yueyang; Fan, Qilin; Yang, Linda
• Is Part Of: Journal of information security and applications. Volume 63(2022)
• Publication Date: 2021-12
• Subjects: Recommender systems -- Shilling attack detection -- Spectral feature -- User similarity; Computer security -- Periodicals; Information technology -- Security measures -- Periodicals; 005.805
• DOI: 10.1016/j.jisa.2021.103051
• Languages: English
• ISSNs: 2214-2126

Abstract

Abstract: Recommender systems can effectively improve user experience, but they are vulnerable to shilling attacks due to their open nature. Attackers inject fake user profiles to destroy the security and reliability of the recommender systems. Therefore, it is crucial to detect shilling attacks effectively. The primitive detection models are feasible but costly because of the dependence on plenty of hand-engineered explicit features based on statistical measures. Even though the upgraded models based on learning embeddings of the implicit features are more general, they fail to take some distinct features in distinguishing fake users into consideration. Moreover, these primitive and upgraded models are difficult to capture the high order relationships between users and items as the models usually learn the embedding from the first-order interactions. The representation and similarity information learned from the first-order interactions are not comprehensive enough, limiting the detection task. To this end, we propose a novel shilling attack detection model by fusing hypergraph spectral features (SpDetector). The proposed model combines the explicit and implicit features to balance the effectiveness and generality and deal with the high order relationships by hypergraphs-based embedding. From the implicit perspective, SpDetector constructs user hypergraphs and item hypergraphs for the high-order relationships hidden in the interaction and extracts spectral features from hypergraphs to capture high-order similarity for users and items, respectively. From the explicit perspective, it extracts two kinds of explicit features: item similarity offsets (ISO) based on item spectral features and rating prediction errors (RPE), for all users as their distinct capability of distinguishing fake users. Finally, the SpDetector learns to distinguish fake users by training a deep neural network with those features. Experiments conducted on MovieLens and Amazon datasets show that SpDetector outperforms state-of-the-art detection models.