AFRL: Adaptive federated reinforcement learning for intelligent jamming defense in FANET

Abstract

The flying ad-hoc network (FANET) is a decentralized communication network for the unmanned aerial vehicles (UAVs). Because of the wireless nature and the unique network properties, FANET remains vulnerable to jamming attack with additional challenges. First, a decision from a centralized knowledge base is unsuitable because of the communication and power constraints in FANET. Second, the high mobility and the low density of the UAVs in FANET require constant adaptation to newly explored spatial environments containing unbalanced data; rendering a distributed jamming detection mechanism inadequate. Third, taking model-based jamming defense actions in a newly explored environment, without a precise estimation of the transitional probabilities, is challenging. Therefore, we propose an adaptive federated reinforcement learning-based jamming attack defense strategy. We developed a model-free Q-learning mechanism with an adaptive exploration-exploitation epsilon-greedy policy, directed by an ondevice federated jamming detection mechanism. The simulation results revealed that the proposed adaptive federated reinforcement learning-based defense strategy outperformed the baseline methods by significantly reducing the number of en route jammer location hop counts. The results also showed that the average accuracy of the federated jamming detection mechanism, leveraged in the defense strategy, was 39.9% higher than that of the distributed mechanism verified with the standard CRAWDAD jamming attack dataset and the ns-3 simulated FANET jamming attack dataset.