Research On Shilling Attack Detection Method Based On Time Series

Nowadays,the development in science and technology made increasing data on the Internet,which has caused trouble for users to choose a lot of information.However,the recommendation system can solve this difficult problem.Faced with a large amount of data,it can help users choose the information that they are interested in and implement personalized services according to user preferences.However,with the widespread application of recommendation systems,traditional collaborative filtering based recommendation present some security vulnerabilities because they rely heavily on user behavior data.Some malicious businesses artificially control the recommendation results by shilling attack.This paper conducts research on shilling attack,and proposes a detecting shilling attack method based on time series.The main research results in this paper are as follows:(1)The paper proposes a novel shilling attack detection method based on T-distribution over the Dynamic Time Intervals.Firstly,this method use Dynamic Time Intervals to divide the rating history of items into multiple time windows;Secondly,the T-distribution is employed to calculate the similarity between windows;Thirdly,abnormal windows are identified by analyzing the T value,time difference and rating actions of windows;Fourthly,abnormal rating actions are detected by analyzing rating mean.(2)The paper proposes a neural network detection of shilling attack based on user rating history and latent features.This method first analyzes the rating history of user,and proposes Rapidity Score;and employs a multilayer perceptron and a generalized matrix decomposition to learn the potential features between the user and the item;then,the neural network model is designed to fuse the latent features with the shilling attack features,and learn on the training set,and finally,test on the test set and analyze the performance of the model.(3)Through a large number of experiments,the method in this paper is compared and analyzed with existing related methods.The experiment results testify the superiority of the proposed methods.