DFUN-KDF: Efficient and robust decentralized federated framework for UAV networks via knowledge distillation and filtering

Abstract

Unmanned aerial vehicles (UAVs) are increasingly crucial across various fields. There is a growing interest in using federated learning (FL) methods to enhance the efficiency of UAV operations. Nevertheless, incumbent methods remain encumbered by significant drawbacks, including high energy consumption from extensive parameter exchanges, the imperative for homogeneous networks, and sensitivity to single-point failures. These difficulties are compounded by the unreliable nature of communication channels and the current inability to effectively manage the diversity of UAV models, highlighting the imperative for more resilient and adaptable FL solutions. To address these issues, we propose an efficient and robust decentralized FL framework for heterogeneous UAV networks. Our framework first leverages the knowledge distillation where UAVs transmit embeddings instead of model parameters to reduce the number of transmission parameter. UAVs update their local models using embeddings generated by other UAVs, which also enables UAVs with diverse architectures to participate in training. Moreover, our framework incorporates a filtering mechanism to remove malicious embeddings, ensuring resilience against adversities in UAV networks. Extensive experiments on various datasets validate the effectiveness and practical deployment potential of our framework.