Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

WinFLoRA: Incentivizing Client-Adaptive Aggregation in Federated LoRA under Privacy Heterogeneity

Mengsha Kou, Xiaoyu Xia, Ziqi Wang, Ibrahim Khalil, Runkun Luo, Jingwen Zhou, Minhui Xue

TL;DR

WinFLoRA addresses privacy heterogeneity in federated LoRA fine-tuning for LLM-powered services by introducing a noise-aware aggregation scheme. The server estimates per-client DP-induced noise via LO O-PCA on uploaded adapters and assigns aggregation weights inversely proportional to the noise, $\,w_i^t \propto 1/(\\hat{\\sigma}_i^t + \tau)$, thereby up-weighting higher-quality updates and down-weighting noisier ones. Clients also adapt their privacy investments using a UCB-based strategy over a finite noise set, ensuring long-term utility alignment with the system objective within a stochastic aggregative Markov game framework, which is shown to admit a stationary Markov equilibrium. Empirically, WinFLoRA yields up to 52.58% higher global accuracy and up to 2.56x higher average client utility compared to state-of-the-art benchmarks, across multiple models and datasets, validating the practicality and scalability of noise-aware aggregation without third-party incentives. This approach enables privacy-preserving, decentralized fine-tuning of large language models with robust performance gains and aligned client incentives.

Abstract

Large Language Models (LLMs) increasingly underpin intelligent web applications, from chatbots to search and recommendation, where efficient specialization is essential. Low-Rank Adaptation (LoRA) enables such adaptation with minimal overhead, while federated LoRA allows web service providers to fine-tune shared models without data sharing. However, in privacy-sensitive deployments, clients inject varying levels of differential privacy (DP) noise, creating privacy heterogeneity that misaligns individual incentives and global performance. In this paper, we propose WinFLoRA, a privacy-heterogeneous federated LoRA that utilizes aggregation weights as incentives with noise awareness. Specifically, the noises from clients are estimated based on the uploaded LoRA adapters. A larger weight indicates greater influence on the global model and better downstream task performance, rewarding lower-noise contributions. By up-weighting low-noise updates, WinFLoRA improves global accuracy while accommodating clients' heterogeneous privacy requirements. Consequently, WinFLoRA aligns heterogeneous client utility in terms of privacy and downstream performance with global model objectives without third-party involvement. Extensive evaluations demonstrate that across multiple LLMs and datasets, WinFLoRA achieves up to 52.58% higher global accuracy and up to 2.56x client utility than state-of-the-art benchmarks. Source code is publicly available at https://github.com/koums24/WinFLoRA.git.

WinFLoRA: Incentivizing Client-Adaptive Aggregation in Federated LoRA under Privacy Heterogeneity

TL;DR

WinFLoRA addresses privacy heterogeneity in federated LoRA fine-tuning for LLM-powered services by introducing a noise-aware aggregation scheme. The server estimates per-client DP-induced noise via LO O-PCA on uploaded adapters and assigns aggregation weights inversely proportional to the noise, , thereby up-weighting higher-quality updates and down-weighting noisier ones. Clients also adapt their privacy investments using a UCB-based strategy over a finite noise set, ensuring long-term utility alignment with the system objective within a stochastic aggregative Markov game framework, which is shown to admit a stationary Markov equilibrium. Empirically, WinFLoRA yields up to 52.58% higher global accuracy and up to 2.56x higher average client utility compared to state-of-the-art benchmarks, across multiple models and datasets, validating the practicality and scalability of noise-aware aggregation without third-party incentives. This approach enables privacy-preserving, decentralized fine-tuning of large language models with robust performance gains and aligned client incentives.

Abstract

Large Language Models (LLMs) increasingly underpin intelligent web applications, from chatbots to search and recommendation, where efficient specialization is essential. Low-Rank Adaptation (LoRA) enables such adaptation with minimal overhead, while federated LoRA allows web service providers to fine-tune shared models without data sharing. However, in privacy-sensitive deployments, clients inject varying levels of differential privacy (DP) noise, creating privacy heterogeneity that misaligns individual incentives and global performance. In this paper, we propose WinFLoRA, a privacy-heterogeneous federated LoRA that utilizes aggregation weights as incentives with noise awareness. Specifically, the noises from clients are estimated based on the uploaded LoRA adapters. A larger weight indicates greater influence on the global model and better downstream task performance, rewarding lower-noise contributions. By up-weighting low-noise updates, WinFLoRA improves global accuracy while accommodating clients' heterogeneous privacy requirements. Consequently, WinFLoRA aligns heterogeneous client utility in terms of privacy and downstream performance with global model objectives without third-party involvement. Extensive evaluations demonstrate that across multiple LLMs and datasets, WinFLoRA achieves up to 52.58% higher global accuracy and up to 2.56x client utility than state-of-the-art benchmarks. Source code is publicly available at https://github.com/koums24/WinFLoRA.git.
Paper Structure (21 sections, 1 theorem, 19 equations, 11 figures, 2 tables, 3 algorithms)

This paper contains 21 sections, 1 theorem, 19 equations, 11 figures, 2 tables, 3 algorithms.

Table of Contents

  1. Introduction
  2. Preliminaries and Related Work
  3. Preliminaries
  4. Related Work
  5. Problem Formulation
  6. System Model
  7. Problem Statement
  8. System Design
  9. WinFLoRA Overview
  10. Noise-aware Weights Allocation (NWA)
  11. Individual Noise Adaptation (INA)
  12. Theoretical Analysis
  13. Experimental Results
  14. Experimental Setting
  15. Overall Performance
  16. ...and 6 more sections

Key Result

Theorem 1

Assume assumption:1– assumption:5 hold. Then there exists a stationary Markov equilibrium $\pi^*$ satisfying eq:equilibrium for all $c_i\in\mathcal{C}$.

Figures (11)

  • Figure 1: Clients apply different levels of noise based on data sensitivity, with higher-sensitivity clients adding stronger perturbations before uploading updates over time.
  • Figure 2: System overview of WinFLoRA. In each round, clients train their own LoRA with newly arrived local data, inject client-specific noise, and upload perturbed adapters. The server estimates per-client noise and assigns weights inversely, then aggregates the weighted updates to form a global fine-tuned model.
  • Figure 3: Example of LOO–PCA residual noise estimation. This figure illustrates the per-client noise estimation procedure on LoRA $B$. A larger residual $||r_i||_2^2$ indicates a larger estimated noise.
  • Figure 4: Performance on AGNews with and without NWA. “Without NWA” refers to a no-incentive baseline with average aggregation. Clients adapt noise with INA in both settings.
  • Figure 5: Performance vs. noise action scales. $\mathcal{A}_G$ and $\overline{\mathcal{U}_L}$ on the AGNews dataset with WinFLoRA under two optional noise scale sets, compared against a w/o NWA benchmark. Top: moderate set $\Sigma_{mod}=\{0.0,0.2,\cdots,1.0\}$. Bottom: fine set $\Sigma_{fine}=\{0.0,0.1,\cdots,1.0\}$.
  • ...and 6 more figures

Theorems & Definitions (3)

  • Definition 1: Stationary Markov equilibrium
  • Theorem 1: Existence of stationary Markov equilibrium
  • proof