A Rotating Server Scheme for Secure Federated Learning in Networked Autonomous Driving

Edge intelligence and federated learning (FL), as key enablers of 6G, is a promising solution for networked Autonomous Driving (NAD). However, traditional federated learning is a server-client architecture, which makes the model training overly dependent on a fixed single aggregation server and makes the FL process insecure and unreliable due to the vulnerability of the aggregation server to a single point of failure. In this paper, we propose a rotating server FL scheme (RSFL) to solve the problem of single point of failure and limited resources and improve environmental adaptability. Specifically, we consider multiple factors to measure the vehicle performance and find the vehicle with the highest performance score in this round as the server for the next round while setting weights that are randomized in each round, which reduces even more the likelihood that a malicious user will recognize the regularity of the chosen server. Finally, the performance of RSFL is evaluated through a large number of experiments, and the results show that compared with baseline FL, FL with randomly selected servers, and peer-to-peer decentralized FL, RSFL can effectively reduce the cases of servers being detected and attacked by malicious adversaries, and improve the accuracy of the model.Edge intelligence and federated learning (FL), as key enablers of 6G, is a promising solution for networked Autonomous Driving (NAD). However, traditional federated learning is a server-client architecture, which makes the model training overly dependent on a fixed single aggregation server and makes the FL process insecure and unreliable due to the vulnerability of the aggregation server to a single point of failure. In this paper, we propose a rotating server FL scheme (RSFL) to solve the problem of single point of failure and limited resources and improve environmental adaptability. Specifically, we consider multiple factors to measure the vehicle performance and find the vehicle with the highest performance score in this round as the server for the next round while setting weights that are randomized in each round, which reduces even more the likelihood that a malicious user will recognize the regularity of the chosen server. Finally, the performance of RSFL is evaluated through a large number of experiments, and the results show that compared with baseline FL, FL with randomly selected servers, and peer-to-peer decentralized FL, RSFL can effectively reduce the cases of servers being detected and attacked by malicious adversaries, and improve the accuracy of the model.