Research On Shilling Attack Detection Method In Recommender Systems Based On Variational Autoencoder And Supervised Prototype Network

With the rapid development of the Internet,recommendation systems have become a core component of various e-commerce and social media platforms.However,recommendation systems also face some security issues,including shilling attacks.Shilling attack refers to malicious users influencing the recommendation results and evaluations of a recommendation system by simulating multiple false accounts or providing false evaluations,in order to gain their own benefits.The existence of shilling attack behavior not only affects users’ shopping experience,but also disrupts the normal operation of recommendation systems,affecting the economic benefits of the platform.Therefore,many experts and scholars are exploring effective methods to solve this problem.Some researchers attempt to identify attack behavior by analyzing user behavior characteristics,including traditional machine learning models and deep learning based models.Although deep neural network models have made some progress in detecting shilling attack in recommendation systems,there are still some challenges.Firstly,in the recommendation system,the number of attacking users is much lower than the number of normal users,which is known as category imbalance,which can lead to significant errors in the model’s processing of attack data;Secondly,attackers can use multiple methods to simulate different types of attacks,and the diversity of attack categories makes it difficult for the model to effectively detect all types of attacks;Finally,the current methods do not address the issue of how to learn a set of manually constructed features,such as variational autoencoders,which cannot accurately distinguish between normal users and attacking users when learning manually constructed features,especially when the number of features is small.In order to address the current problems in neural network models,such as imbalanced data between normal and attacking users,insufficient generalization ability due to multiple types of attacks,and poor detection performance when learning manual feature construction,this paper first proposes a supervised prototype Variational autoencoder(SP-VAE)based attack detection algorithm.This method takes into account the problem of poor detection performance of variational self coding methods when learning manually constructed features,and uses a combination of variational self coding embedding module and iterative prototype classification module method to solve the learning problem of manually constructed features.At the same time,this method combines the advantages of prototype network methods in dealing with small sample data and multi classification problems,solving the problem of data imbalance and the detection of multiple shilling attacks.This study verified that the SP-VAE algorithm outperforms the comparison algorithm in detecting cold start and small-scale shilling attacks on real datasets.However,after further in-depth research,it can be found that the SP-VAE algorithm has some shortcomings,such as a fixed number of prototypes,a fixed distribution of prototypes,sensitivity to noisy data,and inapplicability to large-scale data.These shortcomings limit the expression ability of SP-VAE when dealing with complex data distributions,and reduce the accuracy of detection when facing larger datasets.To address this issue,this paper proposes an attack detection algorithm based on infinite mixed prototype variational autoencoder(IMP-VAE).This method uses a prototype network method based on an infinite hybrid model to replace the previous supervised prototype network module,so that the IMP-VAE algorithm maintains good detection performance in the face of large attack scale and high attack fill rate shilling attack.This study also verified the superiority of this algorithm over traditional prototype network methods,SP-VAE algorithm,and other comparative algorithms on real datasets.