| In recent years,the algorithm used in the recommendation system has gradually changed from the traditional algorithm based on collaborative filtering to the direction of deep learning.The adversarial nature of deep learning models makes deep learning-based recommendation systems more vulnerable to attack,while the effectiveness of deep learning models makes attackers enhance attack algorithms.Typically,the attacker poisons the recommendation system based on deep learning by using a generative adversarial network model to simulate the rating distribution of real users to make a user attack overview,resulting in a degrading performance of the recommendation system.The poisoning attack problem in the recommendation system seriously affects the performance of the recommendation system,resulting in damage to the interests of the platform and an unbalanced user experience.The existing detection algorithm does not effectively extract static extreme features in the user and project matrix,resulting in poor user embedding representation.Moreover,the existing detection algorithm does not effectively extract the dynamic correlation features of user items in the user and project matrix,resulting in low detection accuracy.In order to solve the above problems,this paper proposes two detection algorithms to detect poisoning attacks in recommender systems.Firstly,aiming at the problem that the existing detection methods do not effectively extract static extreme features in the user and item matrix,resulting in poor representation of node embeddings,this paper proposes a poisoning attack detection algorithm based on graph embedding.The algorithm constructs an LMGCN model to extract first-order neighborhood extreme features,constructs a meta-path similarity matrix to extract second-order neighborhood extreme features,and uses a decision tree algorithm to classify the fusion features to complete the poisoning attack detection task.Secondly,aiming at the problem that the existing detection methods do not effectively use the dynamic correlation features in the user and project matrix,resulting in low detection accuracy,this paper proposes a poisoning attack detection algorithm based on reinforcement learning.The algorithm designs reinforcement learning models through user item interaction matrix,user item scoring matrix,user relationship matrix,and user item relationship matrix,solves the reinforcement model to obtain multiple user template sequences,and uses the K-means clustering algorithm to cluster with the user template sequence as the initial clustering center,and identifies the fake user cluster set to complete the poisoning attack detection task.Finally,the two detection algorithms proposed in this paper are experimented on different data sets in the face of different attack algorithms,different attack scales,and verify the effectiveness of the algorithms. |
