Diffusion model for efficient federated learning considering non-IID data in industry scenarios

Guo, Wei, Sang, Hongqiang, Liu, Ying ORCID: https://orcid.org/0000-0001-9319-5940, Li, Qingzong and Cai, Kai 2026. Diffusion model for efficient federated learning considering non-IID data in industry scenarios. Applied Soft Computing 196 , 115120. 10.1016/j.asoc.2026.115120

PDF - Accepted Post-Print Version Available under License Creative Commons Attribution. Download (1MB)

Abstract

Federated Learning (FL) faces challenges due to non-independent and identically distributed (non-IID) data, global model initialization (cold-start), and inefficient model training, especially for industrial FL (IFL). This paper proposes an efficient FL framework incorporating a Conditional Diffusion Model (CDM) and a Non-dominated Sorting Genetic Algorithm II (NSGA-II) to address non-IID data issues and optimize IFL training efficiency. First, clients generate synthetic data using CDM based on a data generation plan provided by the server, ensuring data privacy while improving data distribution and volume. The generated data is utilised for global model initialization and local dataset enhancement for clients. Then, NSGA-II is applied to optimize the volume of training data and the number of local training epochs, balancing computational cost and model accuracy for efficient IFL training. A case study on steel plate surface defect detection validates the proposed method. Results show that global model initialization using CDM significantly improves initial model accuracy compared to traditional methods. The optimized training strategy, implemented via NSGA-II, enhances IFL efficiency by reducing training time while maintaining high model performance. After only 4 global training rounds, the proposed method achieves an accuracy of 0.956, whereas random initialization requires about 20 rounds to reach comparable performance (0.949), corresponding to an 80% reduction in communication rounds. It also improves accuracy over cGAN/cVAE by +5-10 percentage points (e.g., Acc@R0: 0.81 vs 0.76/0.71; Acc@R3: 0.95 vs 0.90/0.86). Holm-corrected paired tests confirm statistical significance (e.g., p = 0.0002 at R0), indicating robust and practically meaningful gains for IFL.

Item Type:	Article
Date Type:	Publication
Status:	Published
Schools:	Schools > Engineering
Additional Information:	RRS policy applied
Publisher:	Elsevier
ISSN:	1568-4946
Date of First Compliant Deposit:	1 April 2026
Date of Acceptance:	25 March 2026
Last Modified:	06 May 2026 11:15
URI:	https://orca.cardiff.ac.uk/id/eprint/186125

Actions (repository staff only)

Edit Item