Eliminating Stealthy Backdoors via Occupying and Withdrawing for Secure Personalized Federated Learning in UAV Swarms

Bo Wang, Zan Zhou, Shujie Yang, Ziwei Li, Xiangliang Ma, Changqiao Xu

Abstract: The rapid emergence of the Low-Altitude Economy has driven the large-scale deployment of uncrewed aerial vehicle (UAV) swarms, where Personalized Federated Learning (PFL) serves as a promising paradigm to enable intelligent and privacy-preserving model training across heterogeneous UAVs. However, the decentralized structure of PFL also exposes UAV swarms to stealthy backdoor attacks, which can manipulate model predictions without degrading main-task accuracy, thereby posing severe safety risks to UAV swarms. To address this challenge, this paper proposes PFL-OW, a novel backdoor elimination mothod grounded in the Occupying and Withdrawing mechanism. The core idea of PFL-OW is to exploit the inherent correlation between backdoor attacks targeting the same class. In Occupying phase, each benign UAV constructs class-wise shadow datasets and injects deliberate backdoors to observe update divergence, through which suspicious target classes are identified using Local Outlier Factor. In Withdrawing phase, reversed learning is performed on these classes with shadow dataset to eliminate stealthy backdoors. Extensive experiments on MNIST and FashionMNIST under five representative PFL frameworks and five stealthy backdoor attacks demonstrate that PFL-OW achieves superior defensive performance in attack success rate while maintaining competitive main-task accuracy. This work ensures secure PFL in UAV swarms against backdoor attacks, thereby reinforcing the reliability and safety foundations of the emerging Low-Altitude Economy.

External IDs:dblp:journals/tccn/WangZYLMX26