DATA-FedAVG: Delay-Aware Truncated Accuracy-Based Federated Averaging for Intrusion Detection in UAV Network

초록

In federated learning, the synchronous approach employed by the aggregating algorithm in the federal server, such as federated averaging (FedAVG), introduces high network communication costs, thus rendering it unsuitable for securing a network of unmanned aerial vehicles. This approach impedes the convergence speed of the global model and degrades its performance by increasing the number of participants. This study proposes a novel optimized aggregating algorithm called delay-aware truncated accuracy-(DATA)-based FedAVG. DATA-FedAVG is robust to the contingencies of straggling edge servers/clients (owing to network connectivity issues and system heterogeneity) and adaptively selects the fraction of clients whose model parameters are to be utilized in building the global model, thus optimally detecting intrusions in the network. In addition, the truncated client selection mechanism applied by DATA-FedAVG allows only clients with high-accuracy contributions to participate in both local training and federal updates. Extensive simulation experiments performed with a cybersecurity dataset validate the high performance of the proposed model and its reliability in accurately detecting attacks within an almost 75% reduced communication cost, while improving the performance of the intrusion detection model in terms of the average accuracy, recall, precision, and F1-score by 2%, 3%, 3%, and 3%, respectively.

키워드