Heterogeneity-Aware Federated Optimization with Secure Aggregation for Distributed Data Mining

Abstract

This paper addresses the challenges of collaborative modeling in distributed data mining scenarios, stemming from difficulties in data centralization, strict privacy constraints, and significant client heterogeneity. It proposes an intelligent modeling framework integrating federated learning. This method uses communication rounds as its core organizational structure, enabling parameter-only collaborative training between a central server and multiple edge nodes. Clients define empirical risks based on local data and perform gradient updates, using regularization terms and proximal constraints to suppress local drift and improve optimization stability. The server performs weighted aggregation based on client sample proportions and the participating set, combining update increment metrics and pruning mechanisms to limit the impact of anomalous updates on the global model. Simultaneously, secure aggregation and update perturbation strategies are introduced to enhance the privacy protection and controllable leakage risk of uploaded information without exposing the original data. The method has a clear, modular, and composable overall structure, adaptable to varying client participation sizes and communication constraints, while balancing prediction performance and probabilistic reliability. Experiments are conducted under a unified evaluation protocol, comparing the proposed framework with several representative methods to verify its effectiveness and applicability in distributed datamining tasks.