Federal learning backdoor attack method for graph data

The invention discloses a federated learning backdoor attack method for graph data, and the method comprises the steps: firstly building a federated learning training system, initializing a client, and randomly selecting a malicious client; secondly, the malicious client generates an adaptive trigger sub-graph corresponding to local data according to the local graph data, and the optimal adaptive trigger sub-graph is solved and embedded into the local data of the malicious client; then, all local clients carry out training according to the initial global model, local model parameters are calculated and uploaded to a central server, and the central server carries out weighted average on all the local model parameters and updates global model parameters; and finally, the central server issues the updated model parameters to each local client, and all the local clients update the model parameters. According to the method, the success rate of graph federation backdoor attacks can be improved while the precision o