FedWGCA: a federated learning based AAV intrusion detection with gradient clipping and attention-based neural networks

Abstract

Autonomous Aerial Vehicles (AAVs) are progressively employed in various applications, including surveillance and logistics. However, their rising usage is coupled by increased cybersecurity threats. Conventional intrusion detection systems (IDS) frequently inadequately address specific problems presented by AAV networks, including dynamic operational environments, varied data distributions, and severe resource constraints. Federated Learning (FL) has emerged as a viable alternative, offering a decentralized approach to collaborative training of intrusion detection models while preserving data privacy. However, FL is not without its shortcomings, including poisoning and backdoor attacks, which can impair the accuracy and reliability of the models, thereby exposing AAVs to advanced cyber threats. This research attempts to design resilient FL-based frameworks that address these drawbacks, boosting the security and resilience of AAV networks in the face of emerging cyber hazards. We present our proposed weighted gradient clipping aggregation (FedWGCA) framework to mitigate the impact of malicious updates during the model training process. Experimental studies show that our FedWGCA outperforms state-of-the-art methods, surpassing FedAvg, FedNova, FedOpt, and FedProx by up to 7.23% in accuracy, 9.99% in precision, and 10.14% in recall. For enhancing resilience in our FL architecture, we further present our robust attention neural network, ANET, which outperforms XGBoost by 0.90% and DNN by 7.10%, showcasing its superior precision and reduced false positives in local training.