Communication-Efficient Improved Federated Learning Framework with Local Privacy Perturbation for Distributed Sensitive Data Privacy Preservation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Federated Learning (FL) emerges as a promising distributed machine learning paradigm, which enables collaborative model training among multiple edge clients without uploading raw sensitive data to the central server. Nevertheless, conventional federated learning systems still suffer from two intractable bottlenecks: severe privacy leakage risks caused by plaintext gradient/model parameter transmission and excessive communication overhead induced by frequent complete parameter updates across heterogeneous edge devices. To tackle the above dual challenges, this paper proposes a communication-efficient improved federated learning framework with adaptive local privacy perturbation, named Fed-LPCC (Federated Learning with Local Privacy Perturbation and Compressed Communication). First, an adaptive mixed Gaussian-Laplace local differential privacy (LDP) perturbation mechanism is designed to inject differentiated noise into local gradient updates, which dynamically adjusts noise amplitude according to gradient sensitivity and residual privacy budget, effectively preventing sensitive data leakage from gradient inversion attacks while reducing unnecessary utility loss. Second, we propose a dual-layer gradient sparsification-quantization compression strategy to eliminate redundant parameters in local updates combined with adaptive client sampling, the framework significantly cuts down uplink and downlink communication traffic during global aggregation. Furthermore, we derive rigorous privacy accounting based on Rényi Differential Privacy (RDP) and complete convergence proof for the proposed Fed-LPCC framework. Extensive comparative experiments are conducted on three mainstream public datasets (MNIST, CIFAR-10, Adult) against state-of-the-art privacy-preserving federated learning methods. The experimental results demonstrate that Fed-LPCC achieves a maximum communication compression ratio of 94.7%, reduces privacy budget consumption by 18.3% compared with traditional fixed-noise LDP-FL methods, and maintains competitive model accuracy with a maximum accuracy drop of only 2.1%. The proposed framework can well balance privacy preservation, communication efficiency and model utility, which is suitable for distributed sensitive data training scenarios in intelligent edge computing, healthcare and financial risk prediction.