Anomaly detection-based attention purification graph defense method

An anomaly detection-based attention purification graph defense method, comprising:

• S1:
• constructing and training an anomaly detector
• wherein the anomaly detector uses a graph link prediction model,
• training the anomaly detector comprises:
• passing graph data into the graph link prediction model for training to obtain an anomaly detector AD, wherein the graph data comprises an undirected unweighted graph =(, ε), wherein an adjacency matrix is A, a degree matrix is D and a feature matrix is N×D matrix X;
• wherein represents the undirected unweighted graph, represents a set of nodes, ε represents a set of edges, and N is a number of nodes of an input graph;
• S2:
• performing global prediction on the graph data by using the trained anomaly detector AD to obtain a linkpred matrix;
• S3:
• inputting the linkpred matrix into a sigmoid function for normalization, and scaling element values of the linkpred matrix to an interval of [0, 1] as detection probability;
• S4:
• training an Anomaly Detection-based Graph Attention Network (ADGAT) model
• S4-1:
• first, preprocessing the feature matrix X and the adjacency matrix A of a graph, unifying X and A as sparse matrices, normalizing X in rows, normalizing A symmetrically, thereafter, standardizing the values of X and A, finally, converting formats of X and A from the sparse matrices into tensors which are named as feature and adj parameters, respectively, and then passing the linkpred matrix, feature and adj into the ADGAT model;
• S4-2:
• in training process, according to the detection probability in the linkpred matrix, dividing three edge clusters {Edge1, Edge2, Edge3} as a low confidence edge set, a common confidence edge set and a high confidence edge set respectively, and performing an attention purification operation on different edge clusters to obtain a purified attention coefficient;
• S4-3:
• passing a message based on the purified attention coefficient, and obtaining a final hidden layer feature through iteration;
• and
• S5:
• using the final hidden layer feature to perform a graph node classification for assigning graph nodes of unknown categories in the graph to a predefined category.