Research On Detection Algorithm For Group Shilling Attacks Based On Tensor In Recommender Systems

The recommendation system has been successfully applied in e-commerce and other fields,effectively solving the problem of information overload in the network.However,in order to achieve profitability,some merchants would inject a large number of false user profiles into the recommendation system to change the probability of the target product being recommended.And this behavior undermines the fairness of the recommendation system.In recent years,shilling attacks in recommendation systems have gradually developed into collusive group shilling attacks.Compared with traditional attack methods,group shilling attacks have the characteristics of strong concealment and offensiveness,which seriously affects the accuracy of the recommended results.Therefore,how to accurately detect shilling group is a key issue in the current recommendation system field.In response to the above-mentioned problems,this article conducts an in-depth study from the essential characteristics of group shilling attacks.First of all,aiming at the problem of low detection accuracy of existing group shilling attacks detection methods,this paper comprehensively considers the characteristics of the shilling attack group’s rating time concentration and the rating similarity of the target item,and proposes a group attack detection algorithm based on the maximum density subtensor mining.The algorithm divides the time window of the rating time series of each item,and establishes a three-dimensional tensor data model to form item tensor groups.Use the M-Zoom model to mine the maximum density subtensor of each item,and filter out the subtensor users with higher behavioral consistency as the candidate group.Extract the shilling attack group characteristics,and use the clustering method to get the attack group.Secondly,aiming at the problem that existing group shilling attack detection methods mainly rely on artificial features to detect attack users,this paper designs a double input convolutional neural network model as a feature extractor to automatically extract the user’s deep features.On the basis of constructing the item popularity rating matrix and relevance matrix of each user,the algorithm extracts the multi-scale features of the two matrices through the deep learning model fusing the convolutional neural network and the feature pyramid network.Combine two multi-scale features to obtain fusion features,and classify them according to the fusion features to obtain group shilling attackers.Finally,the experimental results of the two detection algorithms on Netflix and Amazon datasets are analyzed,and the effectiveness of the proposed algorithm is proved by comparing the existing detection methods.