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

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.

💡 Research Summary

**
The paper introduces WinFLoRA, a novel framework for federated fine‑tuning of large language models (LLMs) using Low‑Rank Adaptation (LoRA) while explicitly handling privacy heterogeneity—the situation where different clients inject different amounts of differential‑privacy (DP) noise according to their data sensitivity or regulatory constraints. Traditional federated learning (FL) assumes either uniform privacy budgets or simply discards noisy updates, which leads to two major problems: (1) high‑noise clients can degrade the global model, and (2) low‑noise clients, which could provide the most useful updates, receive no explicit incentive to keep their noise low, potentially causing them to increase privacy noise unnecessarily.

WinFLoRA addresses these issues through two core mechanisms: Noise‑aware Weight Allocation (NWA) and Client Utility Alignment.

1. Noise‑aware Weight Allocation (NWA)

   • Noise Estimation: After each client uploads its LoRA adapter (a set of low‑dimensional weight matrices), the server estimates the variance contributed by the injected Gaussian DP noise. By comparing the observed variance of the uploaded parameters with the expected variance of a clean LoRA update (derived from the prior model), the server computes an estimate (\hat{\sigma}_k) for each client (k).
   • Weight Computation: The inverse of the estimated noise standard deviation, (w_k = 1/\hat{\sigma}_k), is normalized across all participating clients to produce aggregation weights (\alpha_k). Consequently, clients that add less noise receive larger influence in the weighted average, while high‑noise clients are down‑weighted. This extends the classic FedAvg scheme, which treats all clients equally, by making the aggregation directly proportional to the quality of each update.

2. Client Utility Alignment

   • Each client selects a privacy budget (\epsilon_k) (larger (\epsilon) means less noise). The client’s utility is modeled as a trade‑off between performance gain and privacy loss:
     \