Detecting Fake Co-visitation Injection Attack in Graph-based Recommendation Systems

Recommendation systems are vulnerable to injection attacks by malicious users due to their fundamental openness. One of the vulnerabilities is the fake co-visitation injection attack, which significantly impacts recommendation systems since it modifies the system according to the attacker’s wishes. To date, the detection of co-visitation injection attacks is challenging as: (1) the choice of attribute representation of nodes is hard, (2) practical evidence for analyzing and detecting anomalies in real-world data is insufficient, (3) it is challenging to filter between the original and injected co-visitation data in terms of node behaviors. This paper investigates a unified detection framework that combines attribute and network structure information synergistically to detect outlier nodes based on CUR decomposition and residual analysis. At first, co-visitation graphs are constructed using association rules, and attribute representations of their nodes are developed. Then, both attributes and network structure information are blended in order to identify suspicious nodes. Extensive experiments on both synthetic and real-world dataset exhibit the efficacy of the proposed detection approach compared to other state-of-the-art approaches. The experimental results show that the detection performance can improve by up to 50% for co-visitation injection attacks over the baselines in terms of false alarm rate (FAR) while keeping the highest detection rate (DR).