Secure aggregation for semi-decentralized federated learning under heterogeneous data
Semi-decentralized federated learning has gained attention for combining the advantages of centralized and decentralized approaches, thereby enhancing system scalability and flexibility. However, in the open heterogeneous federated learning architecture, the existence of semi-honest participants and...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
POSTS&TELECOM PRESS Co., LTD
2025-06-01
|
| Series: | 网络与信息安全学报 |
| Subjects: | |
| Online Access: | http://www.cjnis.com.cn/thesisDetails#10.11959/j.issn.2096-109x.2025029 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Semi-decentralized federated learning has gained attention for combining the advantages of centralized and decentralized approaches, thereby enhancing system scalability and flexibility. However, in the open heterogeneous federated learning architecture, the existence of semi-honest participants and heterogeneous data has aggravated the problems of node privacy leakage and difficult model convergence during model aggregation. The differences between local models have increased with the degree of local data heterogeneity, which has further slowed down the global model’ convergence and reduced its accuracy. Additionally, the complex network topology among nodes has heightened the risk of local data privacy leakage, making it difficult to ensure local data privacy protection. To tackle these issues, a secure aggregation scheme for heterogeneous data under a semi-distributed federated learning architecture was designed. This scheme aimed to enhance the global model’s convergence performance while safeguarding node privacy. A random masking mechanism was developed based on the alternating direction multiplier method. This mechanism strengthened local data privacy protection and prevented access to individual node models. Moreover, a double-weighted aggregation strategy was proposed. In this strategy, intra-cluster aggregation weights for nodes were determined based on the loss of the global model on different node samples, and global aggregation weights for different clusters were established according to their contribution to the global model. Extensive experiments were carried out on three public standard datasets. The results demonstrate that, compared with advanced schemes, the proposed method improves model convergence speed and accuracy in the context of heterogeneous data. |
|---|---|
| ISSN: | 2096-109X |