What is it about?
This paper proposes DFUN-KDF, an efficient and robust decentralized federated learning framework tailored for UAV networks. Unlike traditional federated learning approaches that rely on centralized servers and frequent model parameter exchanges, DFUN-KDF enables UAVs to collaboratively train models in a fully decentralized manner by transmitting lightweight embeddings generated from public datasets. Through knowledge distillation, UAVs with heterogeneous model architectures can effectively learn from each other while significantly reducing communication overhead and energy consumption. To further enhance robustness in highly dynamic and unreliable UAV environments, the framework incorporates a filtering mechanism that identifies and removes abnormal or malicious embeddings during aggregation. Extensive experiments on EMNIST, CIFAR-10, and the real-world UAV-Human dataset demonstrate that DFUN-KDF achieves strong performance in both static and dynamic network settings, while offering superior robustness, compatibility across heterogeneous models, and substantial communication efficiency.
Featured Image
Photo by Mitch Nielsen on Unsplash
Perspectives
In this work, we aim to address several critical limitations of existing federated learning frameworks when applied to UAV networks, including high communication costs, vulnerability to network instability, and the inability to support heterogeneous model architectures. By combining decentralized federated learning with knowledge distillation, DFUN-KDF provides a practical solution that allows UAVs to collaborate efficiently without relying on a central server. One key strength of our approach is the use of embedding-level communication instead of full model parameters, which greatly reduces energy consumption—an essential factor for resource-constrained UAV systems. Additionally, the proposed filtering mechanism improves system robustness against unreliable links and malicious participants, making the framework more suitable for real-world deployments. We believe this work contributes a scalable and adaptable federated learning paradigm for intelligent UAV networks and opens new directions for applying decentralized learning in dynamic and heterogeneous environments. Future extensions incorporating privacy-enhancing techniques such as differential privacy could further strengthen the applicability of DFUN-KDF in sensitive real-world scenarios.
Gege Jiang
Sun Yat-Sen University
Read the Original
This page is a summary of: DFUN-KDF: Efficient and robust decentralized federated framework for UAV networks via knowledge distillation and filtering, Communications in Transportation Research, December 2025, Tsinghua University Press,
DOI: 10.1016/j.commtr.2025.100173.
You can read the full text:
Contributors
The following have contributed to this page
