Distributed and Secure Federated Learning for Wireless Computing Power Networks

Abstract

The explosively growth of mobile applications imposes much burden on the current computing networks. Wireless Computing Power Network (WCPN), as an emerging computing architecture, can sense and coordinate computing resources through agile wireless communications, and realize distributed intelligence based on federated learning. However, the mobility and heterogeneity of WCPN nodes typically impact the security (e.g., malicious node disturbance) and efficiency of federated learning in WCPN. In light of this, this paper proposes a provable secure and decentralized federated learning based on blockchain for WCPN, where nodes can freely participate or leave the WCPN federated training without authorization and security threats. Particularly, we design a blockchain with proof-of-accuracy (PoAcc) consensus scheme to deeply integrate with the federated learning procedure, in which high-accuracy local models have the priority of aggregation, thus accelerating the convergence of federated learning and improving the efficiency of WCPN. The proposed PoAcc is proved to be secure as long as the ratio of honest to malicious nodes is above a lower bound. To further meet the security requirement of PoAcc, we then propose an evolutionary game-based incentive scheme that incentivizes honest nodes to participate the WCPN federated learning under malicious node disturbance. Numerical results show that the proposed scheme ensures the consistency and security of federated learning in WCPN, while outperforming the benchmarks in terms of model accuracy and resource consumption.