Supervised Prototypical Variational Autoencoder for Shilling Attack Detection in Recommender Systems

Abstract

Collaborative filtering-based recommender systems are vulnerable to shilling attacks. How to detect shilling attacks has become a popular research direction. Some recent works have applied deep learning to the field of shilling attack detection. However, most of the existing deep learning-based shilling attack detection models are based on user-item scoring matrices, which do not apply manual scoring features well and cannot be used to detect cold-start shilling attackers. Thus, we propose a shilling attack detection algorithm based on Supervised Prototypical Variational Auto-Encoder (SP-VAE). Specially, SP-VAE can obtain a unified user-profile representation that can be easily used to down-stream applications of shilling attack detection classifiers. Then, the algorithm constructs the prototype representation of various shilling attacker, and a classifier is used to classify various shilling attack users and normal users. The experimental results show that our method consistently outperforms the traditional method in the case of cold-start profile of the shilling attack.