HDP-FedCD: Data-quality-driven hierarchical federated learning for optimizing privacy protection in non-IID data

Abstract

With the proliferation of Internet of Things (IoT) devices, Federated Learning (FL) has become a key paradigm for collaborative machine learning on decentralized edge data. However, FL remains vulnerable to inference attacks, posing significant privacy concerns, particularly in scenarios with diverse data quality and distribution. Existing privacy protection methods often neglect such heterogeneity, resulting in suboptimal trade-offs between privacy and performance. We propose a Hierarchical Differential Privacy protection scheme in Federated Learning based on Core-Degree (HDP-FedCD), which achieves an optimal balance between privacy and utility by leveraging core-degree as a measure of data quality to dynamically adjust noise levels. Using an adaptive core-degree threshold, HDP-FedCD layers local datasets into core and non-core layers, tailoring noise intensity to data quality: low-intensity noise preserves utility for core-layer data, while high-intensity noise enhances privacy for non-core-layer data. Furthermore, the adaptive threshold mechanism responds to dynamic data distribution changes, ensuring robustness across diverse FL scenarios. Empirical evaluations on image classification tasks demonstrate that HDP-FedCD outperforms state-of-the-art methods in model accuracy and resistance to inference attacks, offering an innovative solution for privacy-preserving federated learning.