Robust Graph Embedding Recommendation Against Data Poisoning Attack

With the development of recommendation system technology, more and more Internet services are applied to recommendation systems. In recommendation systems, matrix factoring is the most widely used technique. However, matrix factoring algorithms are very susceptible to shilling attacks (trust or espionage). The former defends methods against data poisoning attacks focused on detecting individual attack behaviors. But there are few detection methods for group data poisoning attacks. Therefore, we propose a detection method based on Graph Neural Network (GNN) and adversarial learning. We train user-item nodes and edges through a semi-supervised learning approach, improving the robustness of the GNN recommendation system. Our work can be divided into the following parts: Firstly, we review the former recommendation systems and the graph representation learning recommendation systems. Secondly, we analyze the main vulnerabilities of the graph representation learning recommendation systems. Furthermore, the detection methods of data poisoning attacks are analyzed, and the difference between individual data poisoning attacks and group data poisoning attacks are discussed. Finally, we propose a per-process Robust-GNN semi-supervised detection model to conduct group detection on different types of attacks. In addition, we also analyze the sensitivity of the proposed methods. From the experiments results, it can be concluded that we should apply the attention mechanism to the proposed methods which makes it more generalized.