Semi-asynchronous federated learning-based privacy-preserving intrusion detection for advanced metering infrastructure

The emergence of smart grid brings great convenience to users and power companies, but also brings many new problems, among which the most prominent one is network attack security. Although federated learning works well in dealing with smart grid network attacks, it suffers from gradient leakage, client node failure and a single type of training model. Therefore, this paper proposes a semi-asynchronous federated learning-based privacy-preserving intrusion detection for advanced metering infrastructure (AMI). First, we design a hierarchical federated learning framework based on chained secure multiparty computing, which allows concentrators to collaboratively train models to protect local gradients. Second, we adapt the framework to the AMI network structure characteristics, and design a semi-asynchronous model distribution protocol. Finally, we build an ensemble model based on temporal convolutional network and gated recurrent unit (TCN-GRU) to detect AMI network attacks. The experimental results show that the proposed method can achieve 99.23% accuracy than existing methods.