| 基于多轮对话基准测试的大语言模型微调联邦学习激励机制的研究 |
| 点此下载全文 |
| 引用本文：杨梦轲,胡小明,白双杰,刘琰.基于多轮对话基准测试的大语言模型微调联邦学习激励机制的研究[J].上海第二工业大学(中文版),2025,42(4):425-432 |
| 摘要点击次数: 155 |
| 全文下载次数: 15 |
|
|
| 中文摘要:大语言模型(large language models, LLMs) 微调技术近几年受到广泛的研究与应用。目前, 针对LLMs 微调联邦学习的研究集中于工程实践、隐私保护及训练效率等方面。然而, 现有研究大都基于参与联邦学习的节点会无条件配合这一理想假设, 与实际情况往往不符。为解决该问题, 本文提出了一种基于多轮对话基准测试的LLMs 微调联邦学习激励机制。该机制能够公平地为节点贡献分配激励, 促进节点自发地参与联邦学习, 弥补了其在LLMs微调联邦学习场景下应用的空白。同时, 所提激励机制还能防御LLMs 联邦学习微调过程中的投毒攻击, 提升LLMs微调联邦学习的健壮性。最后通过模拟实验验证了该激励机制的有效性及抵御投毒攻击的能力。 |
| 中文关键词:联邦学习 激励机制 大语言模型 大模型微调 低秩适配微调 |
| |
| Research on Federated Learning Incentive Mechanisms for Large Language Models Fine-Tuning Based on Multi-Turn Dialogue Benchmarks |
|
|
| Abstract:Large language models (LLMs) fine-tuning have garnered extensive research and application in recent years. Current research on federated learning for LLMs fine-tuning primarily focuses on engineering practices, privacy protection, and training efficiency. However, current research is predominantly based on the ideal assumption that nodes participating in federated learning will unconditionally engage in the process, which is inconsistent with real-world scenarios. To address this issue, this paper proposes a federated learning incentive mechanism for LLMs fine-tunning based on multi-turn dialogue benchmarks. The proposed mechanism allocates incentives fairly based on node contributions, encouraging voluntary participation in federated learning, filling the gap in the application within the context of federated learning-based fine-tuning of LLMs. Meanwhile, the incentive mechanism can also defend against poisoning attacks during LLMs federated learning fine-tunning, enhancing the robustness of the LLMs fine-tuning process. Finally, through simulation experiments, the effectiveness of the proposed incentive mechanism and its effectiveness in mitigating poisoning attacks are validated. |
| keywords:federated learning incentive mechanism large language models large model fine-tuning low-rank adaptation |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|