Hybrid CNN and RNN-based shilling attack framework in social recommender networks

Abstract

INTRODUCTION: Recommender system is considered to be widely utilized in diversified domain for the purpose of effectively handling information overload. But, recommender systems are prone to vulnerabilities that are significantly exploited by malicious attacks. In particular, shilling attack is determined to crucial in the recommender system due to its openness characteristics and data dependence.

OBJECTIVES: Authors focused on detecting shilling attack by using hybrid deep learning techniques.

METHODS: Hybrid CNN and RNNs-based shilling attack framework is proposed for shilling attack detection based on the selection of dynamic features for attaining maximized detection accuracy.

RESULTS: The proposed CNN-RNNs-based shilling attack framework was determined to improve the recall with different filler size under Netflix dataset by 4.48% and 6.14%, better than the benchmarked HDLM and RMRA frameworks. The proposed CNN-RNNs-based shilling attack framework was determined to minimize the false positive rate by 4.82% and 5.94%, better than the benchmarked HDLM and RMRA frameworks.

CONCLUSION: This framework integrated user popularity and rating-based indicators in order to consider the deviations that happens, when the users select items. It also included information entropy for dynamically choosing the detection indicators in order to improve the reliability in attack detection. It was proposed with three different attack detection models that contextually handles different shilling attacks.