结合元学习的联邦检索增强生成模型

Federated Retrieval-augmented Generation Model with Meta-learning

摘要

检索增强生成(RAG)技术在知识密集型问答任务中表现出色,但集中化知识库模式导致数据隐私泄露、传输成本高和实时性不足等问题.现有联邦RAG方法采用静态权重聚合,无法根据查询特征动态调整各客户端贡献度,难以充分利用分布式的专业化优势.本文提出MetaFedRAG框架,旨在实现隐私保护下的智能联邦问答.该框架将元学习机制引入联邦RAG,设计元学习权重预测器,通过学习查询特征与客户端专业能力的映射关系,实现查询感知的动态权重分配.实验结果显示,MetaFedRAG相比基线方法在所有指标上显著提升:损失函数最大降低38.2%,准确率提升0.12%~2.12%,F1分数提升0.37%~7.68%,在异构环境下展现卓越稳定性和鲁棒性.该框架有效缓解数据异构性带来的性能偏差,为隐私敏感应用提供兼顾性能与隐私的创新解决方案.

Abstract

Retrieval-Augmented Generation(RAG) technology performs excellently in knowledge-intensive question-answering tasks,but its centralized knowledge base paradigm leads to data privacy leakage,high transmission costs,and insufficient real-time performance.Existing federated RAG methods adopt static weight aggregation and cannot dynamically adjust each client′s contribution based on query characteristics,making it difficult to fully leverage the specialization advantages in distributed environments.This paper proposes the MetaFedRAG framework,which aims to achieve intelligent federated question-answering with privacy preservation.The framework introduces meta-learning mechanisms into federated RAG and designs a meta-learning weight predictor that learns the mapping relationship between query features and client expertise to achieve query-aware dynamic weight allocation.Experimental results show that MetaFedRAG significantly outperforms baseline methods across all metrics:the loss function decreases by up to 38.2%,accuracy improves by 0.12% to 2.12%,and F1 scores increase by 0.37% to 7.68%,demonstrating exceptional stability and robustness in heterogeneous environments.This framework effectively mitigates performance biases caused by data heterogeneity and provides an innovative solution that balances performance and privacy for privacy-sensitive applications.

关键词

检索增强 / 联邦学习 / 元学习 / 分布式

Key words

retrieval-augmented / federated learning / meta-learning / distributed

[1] Lewis P,Perez E,Piktus A,et al.Retrieval-augmented generation for knowledge-intensive NLP tasks[C]//Advances in Neural Information Processing Systems,2020:9459-9474.
[2] Kairouz P,McMahan H B,Avent B,et al.Advances and open problems in federated learning[J].Foundations and Trends in Machine Learning,2021,14(1-2):1-210.
[3] Addison P,Nguyen M T H,Medan T,et al.C-fedrag:a confidential federated retrieval-augmented generation system[J].arXiv preprint arXiv:2412.13163,2024,doi:10.48550/ arXiv.2412.13163.
[4] Zhao D.Frag:toward federated vector database management for collaborative and secure retrieval-augmented generation[J].arXiv preprint arXiv:2410.13272,2024,doi:10.48550/ arXiv.2410.13272.
[5] McMahan B,Moore E,Ramage D,et al.Communication-efficient learning of deep networks from decentralized data[C]//Artificial Intelligence and Statistics,2017:1273-1282.
[6] Karpukhin V,Oguz B,Min S,et al.Dense passage retrieval for open-domain question answering[C] //Conference on Empirical Methods in Natural Language Processing,2020:6769-6781.
[7] Izacard G,Grave E.Leveraging passage retrieval with generative models for open domain question answering[C]//16th Conference of the European Chapter of the Association for Computational Linguistics,2021:874-880.
[8] Li T,Sahu A K,Zaheer M,et al.Federated optimization in heterogeneous networks[C]//Proceedings of Machine Learning and Systems,2020:429-450.
[9] Karimireddy S P,Kale S,Mohri M,et al.Scaffold:stochastic controlled averaging for federated learning[C]//International Conference on Machine Learning,2020:5132-5143.
[10] Lin Y,Han S,Mao H,et al.Deep gradient compression:reducing the communication bandwidth for distributed training[J].arXiv preprint arXiv:1712.01887,2017,doi:10.48550/ arXiv.1712.01887.
[11] Reisizadeh A,Mokhtari A,Hassani H,et al.Fedpaq:a communication-efficient federated learning method with periodic averaging and quantization[C]//International Conference on Artificial Intelligence and Statistics,2020:2021-2031.
[12] Fallah A,Mokhtari A,Ozdaglar A.Personalized federated learning with theoretical guarantees:a model-agnostic meta-learning approach[C]//Advances in Neural Information Processing Systems,2020:3557-3568.
[13] Dinh T C,Tran N,Nguyen J.Personalized federated learning with moreau envelopes[C]//Advances in Neural Information Processing Systems,2020:21394-21405.
[14] Aashiq M,Thaker P,Diab M T,et al.Cache me if you can:the case for retrieval augmentation in federated learning[C/OL]//Privacy Regulation and Protection in Machine Learning Workshop,https://openreview.net/forum? id= MKd1SkDbbz,2024.
[15] Mao Q,Zhang Q,Hao H,et al.Privacy-preserving federated embedding learning for localized retrieval-augmented generation[J].arXiv preprint arXiv:2504.19101,2025,doi:10.48550/arXiv.2504.19101.
[16] Finn C,Abbeel P,Levine S.Model-agnostic meta-learning for fast adaptation of deep networks[C]//International Conference on Machine Learning,2017:1126-1135.
[17] Chen F,Luo M,Dong Z,et al.Federated meta-learning with fast convergence and efficient communication[J].arXiv.org,2018,doi:10.48550/arXiv.1802.07876.
[18] Hospedales T,Antoniou A,Micaelli P,et al.Meta-learning in neural networks:a survey[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2021,44(9):5149-5169.
[19] Nentidis A,Katsimpra G,Vandoru E,et al.Overview of bioasq 2021:the 9th biosq challenge on large-scale biomedical semantic indexing and question answering[C]//International Conference of the Cross-Language Evaluation forum for European Languages,2021:239-263.

基金资助

国家自然科学基金项目(61903254)资助.