Unsupervised attack method for graph contrast learning

The invention discloses an unsupervised attack method for graph contrast learning, which comprises the following steps: injecting malicious nodes to attack graph data, randomly inserting a plurality of attack nodes on an original graph, and carrying out random connection; and based on graph data after GIA attack, generating a comparison view angle, learning embedding on the same encoder, finally calculating comparison loss, and iteratively attacking an adjacent matrix. According to the method, firstly, a poisoning map is generated by utilizing a comparative learning thought, and information characteristics under different comparative view angles are fully utilized, so that a model can be free from dependence on labels; meanwhile, the GIA method is introduced into the generation process of the poisoning map, and map data are attacked from the aspect of node features. Compared with other methods, the time complexity is lower; and the effect on a plurality of popular data exceeds the SOTA level. 本发明公开了一种针对图对比学习的