Malicious Attack Detection Method for Recommendation Systems Based on Meta-pseudo Labels and Dynamic Features

Abstract

In recommendation systems, malicious users manipulate system recommendations by injecting fake reviews or false co-visit behaviors, thereby gaining undue traffic and exposure, and ultimately compromising the fairness and accuracy of the recommendation results. Existing detection methods largely rely on static analysis of users’ historical ratings and click behaviors, neglecting the dynamic changes in user behavior over time. Additionally, detection methods based on deep learning usually require a large amount of labeled data to achieve excellent detection performance, which is often difficult to meet in practical applications. To address these issues, this paper proposes a malicious attack detection method for recommendation systems based on dynamic features and meta pseudo labels. By constructing a series of dynamic statistical features based on the time series of user behaviors, this method can effectively capture changes in behavior over time. At the same time, the application of meta-pseudo label technology expands the dataset and reduces the dependence on a large amount of labeled data. The use of a soft voting mechanism to integrate the detection results of the meta-pseudo label teacher and student models significantly improves the accuracy and robustness of detection compared to traditional meta-pseudo label methods with a single student model prediction. Experiments on real datasets have verified the high efficiency of this method under the condition of limited labeled samples.