Against network attacks in renewable power plants: Malicious behavior defense for federated learning

As reducing carbon emissions can relieve environmental concerns, networks-supported renewable power plants are being built more and more. Inevitable network attacks have become a serious threat in increasing and distributed power plants. Leveraging federated learning for training the joint model to detect network attacks in distributed power plants is efficient, but two malicious behaviors of cheating and free-riding are unavoidable. To this end, we design a new SDN based federated security architecture and propose a carbon-credit-rewarded consensus verification mechanism in this architecture to deal with malicious behaviors. For this architecture, on the one hand, considering geographical conditions of renewable power plants, multi-controller SDN is adopted in network to solve some security problems at root and to avoid single point of failure. On the other hand, the segmentation of collaborative zones reduces communication cost effectively. The proposed mechanism establishes consensus bearer and realizes the election of consensus bearer by cross-validation of client detection models. Only the excellent models are aggregated to mitigate cheating of malicious clients. Carbon emissions credit is introduced as an incentive in the proposed mechanism. The redistribution of carbon emissions credit improves the performance of global detection model and avoids free-riding. Moreover, the economic nature of carbon emissions credit enhances the spillover effect of carbon emissions trading market on the reduction of carbon emissions. The experimental results revealed that the proposed architecture has excellent performance, and can handle malicious behaviors effectively.