Fine-tuning large language models in federated learning with fairness-aware prompt selection

Area of Science:

• Artificial Intelligence
• Machine Learning
• Natural Language Processing

Background:

• Large language models (LLMs) require domain-specific fine-tuning for deployment.
• Data privacy and computational constraints are critical barriers.
• Federated learning (FL) enables collaborative tuning on private data, preserving confidentiality.

Purpose of the Study:

• To propose FedPSF-LLM, a novel FL framework addressing non-IID degradation, communication overhead, and fairness issues in FL-LLM.
• To enhance privacy-preserving and fairness-guaranteed LLM deployment in federated systems.

Main Methods:

• Prompt Selection Module (PSM) adaptively selects high-impact prompt parameters to reduce transmission costs.
• Dynamic Weighting Module (DWM) adjusts aggregation weights based on client contribution and data disparity.
• Attention-Based Bias Mitigation (ABM) corrects aggregation bias via alignment-aware reweighting.

Main Results:

• FedPSF-LLM improves fairness while maintaining strong overall performance across 10 NLP tasks and 4 LLMs.
• Reduced accuracy variance by 52.1%, improved worst-client accuracy by 8.6%, and narrowed client performance gaps by 74.4%.
• Achieved 76.8% global accuracy, outperforming 8 baselines in fairness and communication efficiency.

Conclusions:

• FedPSF-LLM establishes a new paradigm for privacy-preserving and fairness-guaranteed LLM deployment in federated systems.
• The framework effectively mitigates common challenges in federated learning for large language models.