DEeR: Deviation Eliminating and Noise Regulating for Privacy-preserving Federated Low-rank Adaptation
Meilu Zhu, Axiu Mao, Jun Liu, Yixuan Yuan
TL;DR
This work tackles privacy-preserving federated fine-tuning of foundation models with LoRA, focusing on medical tasks. It identifies aggregation deviation and DP-noise amplification as core challenges when aggregating LoRA updates and introduces DEeR, which combines a server-side deviation eliminator based on generalized alternating minimization and a client-side noise regulator to decouple DP noise from LoRA updates. Theoretical results show that aggregation deviation vanishes when LoRA parameters are equivalent across clients, and empirical results on OCT-C8, Kvasir-v2, M&MS, and Polyp Segmentation demonstrate that DEeR consistently outperforms state-of-the-art LoRA-based FedFT methods under DP constraints. The approach yields robust performance across data heterogeneity, rank choices, communication budgets, and client counts, highlighting its practical potential for privacy-aware medical model finetuning. Code is available at the provided GitHub link.
Abstract
Integrating low-rank adaptation (LoRA) with federated learning (FL) has received widespread attention recently, aiming to adapt pretrained foundation models (FMs) to downstream medical tasks via privacy-preserving decentralized training. However, owing to the direct combination of LoRA and FL, current methods generally undergo two problems, i.e., aggregation deviation, and differential privacy (DP) noise amplification effect. To address these problems, we propose a novel privacy-preserving federated finetuning framework called \underline{D}eviation \underline{E}liminating and Nois\underline{e} \underline{R}egulating (DEeR). Specifically, we firstly theoretically prove that the necessary condition to eliminate aggregation deviation is guaranteing the equivalence between LoRA parameters of clients. Based on the theoretical insight, a deviation eliminator is designed to utilize alternating minimization algorithm to iteratively optimize the zero-initialized and non-zero-initialized parameter matrices of LoRA, ensuring that aggregation deviation always be zeros during training. Furthermore, we also conduct an in-depth analysis of the noise amplification effect and find that this problem is mainly caused by the ``linear relationship'' between DP noise and LoRA parameters. To suppress the noise amplification effect, we propose a noise regulator that exploits two regulator factors to decouple relationship between DP and LoRA, thereby achieving robust privacy protection and excellent finetuning performance. Additionally, we perform comprehensive ablated experiments to verify the effectiveness of the deviation eliminator and noise regulator. DEeR shows better performance on public medical datasets in comparison with state-of-the-art approaches. The code is available at https://github.com/CUHK-AIM-Group/DEeR.