Data Poisoning Attack Detection Methods For Recommendation Systems Based On Time Series And Graph Embeddings

Recommendation systems based on deep learning models,such as neural networks,can extract high-level abstract features from large amounts of sparse and complex data,enabling more accurate and effective recommendation predictions.Recommendation systems are vulnerable to attacks due to their openness and fragility.Attackers can inject false ratings into specific products to manipulate the recommendation results and benefit themselves.This behavior is known as recommendation poisoning attacks.To detect these attacks,various detection methods have been proposed.However,with the continuous improvement of attack models,existing detection methods cannot fully exploit the existing information or represent the high-order relationships in the dataset,resulting in detection performance being limited by the evolution of attack models and unable to distinguish differences between real user profiles and attack user profiles.To address these issues,this paper proposes two recommendation system poisoning attack detection methods based on time series and graph embedding.Firstly,in response to the problem of existing attack detection algorithms failing to fully mine existing information,this paper proposes a poisoning attack detection algorithm based on time series and heterogeneous information networks.This detection algorithm constructs a heterogeneous information network based on user rating behavior,obtains important node sequences through meta-path,and constructs user preference features using the graph2 vec model.It processes the user rating time series using the gramian angular field method and constructs user temporal features using a convolutional neural network.The user preference features and user temporal features are fused and a fully connected network is used for attack user detection.Secondly,in response to the problem of existing detection algorithms failing to capture high-order relationships between users,this paper proposes a poisoning attack detection algorithm based on time series and hypergraph convolutional networks.This detection algorithm uses an LSTM network to perform time series modeling based on user rating times and constructs user temporal feature matrix.It calculates the user correlation based on project time offset and rating offset,constructs a hypergraph based on user correlation,and uses a hypergraph convolutional neural network model to learn high-order relationships and correlation features between nodes.Attack user detection is performed using a fully connected network.Finally,the proposed algorithms are tested on the Movie Lens 1M and Netflix datasets and compared with three existing detection methods.Experiments demonstrate the effectiveness of both detection methods.