Federated Deep Learning for Sustainable Systems: A Privacy-Preserving and Scalable Big Data Analytics Framework
Yuan Zheng, Fu Xiaoyu · 2025
This study proposes a stable and efficient federated deep learning multi-node collaborative optimization framework for big data analysis scenarios. The framework can cope with realistic challenges such as low communication efficiency, strong node heterogeneity, strict resource constraints, and high privacy protection requirements. The framework comprehensively designs four types of optimization mechanisms, namely: communication efficiency optimization, heterogeneous adaptation optimization, resource-aware modeling optimization, and robustness enhancement strategy under privacy constraints. At the communication optimization level, the proposed model can effectively reduce communication rounds and parameter dimensions by introducing communication budget constraints and sparse update mechanisms; at the heterogeneous adaptation level, the proposed model constructs a dynamic local training strategy based on client capability factors. This operation realizes resource-aware scheduling of low-performance nodes; at the resource modeling level, the proposed model incorporates device computing and communication energy consumption into the joint optimization goal. At the same time, the hierarchical reinforcement learning algorithm is used to solve non-convex problems; at the privacy protection level, the proposed model is based on the Gaussian difference privacy mechanism and the Lagrangian dual method. The model completes the theoretical definition of model convergence and the convergence design of the optimization function. The results show that under the premise of ensuring the accuracy of the model, the communication overhead of the proposed algorithm is reduced by an average of about 35%; in a heterogeneous environment with 40% device performance constraints, the average training delay is reduced by about 28%. At the same time, the unit energy consumption accuracy ratio is the highest under resource constraints, and the model accuracy is better than the existing differential privacy algorithm under the same privacy budget conditions.