SAShA: Semantic-Aware Shilling Attacks on Recommender Systems Exploiting Knowledge Graphs

Recommender systems (RS) play a focal position in modern user-centric online services. Among them, collaborative filtering (CF) approaches have shown leading accuracy performance compared to content-based filtering (CBF) methods. Their success is due to an effective exploitation of similarities/correlations encoded in user interaction patterns, which is computed by considering common items users rated in the past. However, their strength is also their weakness. Indeed, a malicious agent can alter recommendations by adding fake user profiles into the platform thereby altering the actual similarity values in an engineered way. The spread of well-curated information available in knowledge graphs (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {KG}$$\end{document}) has opened the door to several new possibilities in compromising the security of a recommender system. In fact, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {KG}$$\end{document} are a wealthy source of information that can dramatically increase the attacker’s (and the defender’s) knowledge of the underlying system. In this paper, we introduce SAShA, a new attack strategy that leverages semantic features extracted from a knowledge graph in order to strengthen the efficacy of the attack to standard CF models. We performed an extensive experimental evaluation in order to investigate whether SAShA is more effective than baseline attacks against CF models by taking into account the impact of various semantic features. Experimental results on two real-world datasets show the usefulness of our strategy in favor of attacker’s capacity in attacking CF models.