Decentralized Federated Learning for Adversarial Anomaly Detection in Consumer-Grade UAV-Assisted MEC Systems

Anomaly detection is a key technique to ensure the security of consumer-grade Unmanned Aerial Vehicle (UAV) assisted Mobile Edge Computing (MEC) systems. While integrating emerging Decentralized Federated Learning (DFL) tends to enhance the efficiency and reduce overhead of anomaly detection, most existing DFL frameworks still face huge challenges in system and statistical heterogeneity. This vulnerability makes them highly susceptible to adversarial attacks during global communication, significantly undermining system robustness. To address these important challenges, we propose <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">DFL2AD</i>, a novel DFL framework for Adversarial Anomaly Detection in consumer-grade UAV-assisted MEC systems. First, we split the local model into top-level and bottom-level components. The bottom-level model is compressed via model compression, and the UAVs only exchange the compressed bottom model with their neighbors, thereby reducing the communication overload of neighboring UAVs. Next, we design a new self-knowledge distillation mechanism. This mechanism uses the local model from the previous round as a teacher model to guide the training of the post-aggregation model in the current round, thereby preventing the forgetting of prior knowledge. Finally, we develop a residual detection strategy for identifying and replacing adversarial models. Notably, we theoretically prove the stable convergence of <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">DFL2AD</i> during the aggregation process. Based on the real-world testbed and traffic datasets, extensive experiments validate the effectiveness of the proposed <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">DFL2AD</i> in solving the statistical and system heterogeneity issues among consumer-grade UAVs. Compared to the state-of-the-art methods, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">DFL2AD</i> achieves higher detection accuracy (by up to around 20%) and training efficiency under adversarial attacks, while exhibiting better scalability and convergence.