Differential Privacy-Based Adaptive Sparse Federated Learning in UAV Networks

With the rapid development of the low-altitude economy, privacy protection has become a significant challenge in the unmanned aerial vehicles (UAV) networks. Federated learning (FL) provides a concrete framework for addressing privacy concerns in the low-altitude networks by enabling training without exposing raw data. However, there remains a risk of data leakage during aggregation of parameter updates from local models in the FL framework. Existing approaches have introduced differential privacy (DP) to mitigate this issue, but adding DP noise can degrade the performance of the training process. To further enhance the efficiency and accuracy of model training, we propose a novel framework based on DP and adaptive sparsity for FL, named DP-FedAS. On the one hand, this framework reduces communication and training overhead through an adaptive sparsity module. On the other hand, it mitigates privacy errors caused by DP noise by reducing the noise introduced during global aggregation via sparsity, thereby alleviating the performance degradation. Furthermore, we provide detailed theoretical proofs for the convergence of the proposed algorithm and the privacy guarantees it offers. Simulation results validate that DP-FedAS improves global model accuracy by 20%, and reduces communication cost by 23%, while maintaining a robust level of privacy protection. The proposed framework strikes an optimal balance among communication efficiency, privacy preservation, and model performance.