PFFLoRA: Personalized Fourier LoRA Fine-Tuning of Federated Large Language Models
Kang Li, Zuoxian Lu, Shilin Yu · 2024
Fine-tuning large language models (LLMs) in federated learning settings presents significant challenges, including substantial communication overhead, non-identically distributed (non-IID) data, and client hardware limitations. Traditional federated learning methods like Federated Averaging (FedAvg) often fail to personalize models for diverse client data, leading to suboptimal performance. We propose Personalized Federated Fourier Low-Rank Adaptation (PFFLoRA), which leverages spectral parameterization to significantly reduce communication costs while enabling model personalization. PFFLoRA updates only a small set of Trainable Fourier coefficients and introduces a personalized contrastive learning loss, allowing clients to fine-tune the global model according to their local data characteristics. Experimental results show that PF-FLoRA reduces communication costs by up to 99.8% compared to full fine-tuning, while achieving performance comparable to full fine-tuning. This offers a scalable and personalized solution for federated learning with large language models.