Online Personalized Federated Learning Methods for Intrusion Detection in Dynamic UAV Networks

Xiaoshan Cui, Junhua Wu, Xin Fan, Qiqi Yu, Tielin Wang, Guangshun Li, Chuanwen Luo

Abstract: Currently, machine learning methods are commonly used for intrusion detection in drones. Centralized machine learning requires collecting data from distributed drones, which can lead to issues such as excessive communication volume, heavy burden on the central node, and easy leakage of privacy. Intrusion detection based on federated learning has attracted attention because it can avoid the privacy risks associated with centralized data transmission. However, most existing federated learning schemes are static and difficult to adapt to the dynamic scenarios of drones, thus limiting their applications. This paper proposes an intrusion detection architecture—Online Personalized Federated Learning (OPFL)—that caters to data dynamics and the heterogeneity of drone devices. The proposed scheme inherits the advantages of federated learning in terms of privacy protection. However, unlike existing federated learning methods, this approach pays more attention to local individual goals. In addition, we apply an adaptive mechanism in local model training, which improves the personalization of each client and the robustness of the global model. Experimental results show that when using this scheme to detect non-independent and identically distributed data in the CICIDS 2017 public dataset, which is widely used in the field of network anomalies, the detection accuracy reaches 95.88%. Compared with APFL and Asynchronous Online Federated Learning, OPFL performs better in detecting attacks.

External IDs:dblp:conf/wasa/CuiWFYWLL25