LAP-DPFL: Loss-Driven Adaptive Pruning for Efficient and Privacy-Preserving Federated Learning

The rapid development of edge computing and privacy requirements has promoted Differentially Private Federated Learning (DP-FL). However, DP-FL still faces critical challenges, including high communication overhead, privacy leakage risks, and performance degradation under non-IID data. Existing methods often rely on static pruning, separate designs of compression and privacy, and poor structure consistency across clients, leading to limited performance. To address these issues, we propose LAP-DPFL, a Loss-Driven Adaptive Pruning framework for Differentially Private Federated Learning. Our method integrates adaptive pruning intensity control, co-optimized sparse differential privacy, and update-history-guided structural consistency synchronization. This design jointly improves communication efficiency, privacy protection, and robustness to non-IID data. Extensive experiments on MNIST, Fashion-MNIST, and CIFAR-10 demonstrate that LAP-DPFL outperforms baseline methods in model accuracy, communication compression, and stability under heterogeneous data distributions.