Detection Method Of Poisoning Attack In Recommender Systems Based On Graph Embedding And Anomaly Detection

In a recommendation system,user rating data provides valuable information to a group of users with similar interests,and the system will be more inclined to recommend products with higher ratings to similar users.However,the organized injection of fake users by black marketers,aiming to increase the probability of target products being recommended or to suppress competitors’ products,have seriously affected the service quality of recommendation systems.For how to detect these fake users,domestic and foreign researchers and scholars have conducted extensive research and achieved good detection results.However,in recent years,with the widespread use of deep learning technology,researchers have used deep learning-related techniques,such as adversarial generative networks,to generate fake users and conduct poisoning attacks against the training set of recommendation systems,which have the characteristics of "small injection scale,large interference effect,and high imitation accuracy".Traditional detection methods often rely on expert knowledge or artificially designed indicators,and cannot make full use of the user-commodity relationship,so the detection effect is poor.In order to solve the above problems,this paper conducts an in-depth study on the detection methods for poisoning attacks on recommender systems.First,we propose a detection method based on multiplexed bipartite graph embedding and deviation networks to address the problem that existing methods rely on expert knowledge or human-designed metrics.The method considers user-goods-rating as a multiway bipartite graph,and uses the rating information to transform it into two sets of isomorphic hypergraphs from different perspectives,which are the set of goods hypergraphs for aggregating information from user perspectives and the set of user hypergraphs for aggregating information from goods perspectives,respectively;then extracts the internal structural features of the set of hypergraphs from different perspectives according to the binding self-encoder,and combines the intra-and inter-domain message passing The homogeneous hypergraph convolution operator with intra-and inter-domain messaging strategy captures semantic and higher-order structural information to obtain a high-quality low-dimensional vector representation of users;the deviation network is used to calculate the suspiciousness value of each user,rank them and detect the poisoning attack users.Second,a detection method based on comparing bipartite graph embedding and deep SAD is proposed to address the problem that existing detection methods ignore the connection between nodes and global information.The method obtains the initial embedding of user nodes and commodity nodes by aggregating the second-order neighbor information of all nodes based on the bipartite graph encoder;obtains the local embedding of the h-order closed subgraph based on the attention mechanism,and obtains the global embedding by combining the simple aggregation function;designs a negative sampling mechanism based on scoring to sample the local negative examples of nodes,and optimizes the model parameters by maximizing the mutual information between the local and global representations to obtain the user embedding vector;after that the sampled feature mapping function projects the user embedding onto the hypersphere,and the samples wrapped by the hypersphere are regarded as real users according to a small amount of label correction,and those deviating from the hypersphere are false users.Finally,the proposed algorithm outperforms the four existing detection methods on four datasets of different sizes and densities,namely Film Trust,ML-100 k,Ciao and Amazon.