| 引用本文： |
-
张慧,张盛兵,罗珍杰,杨黎斌,慕德俊,马建峰.联邦大模型研究现状及应用挑战综述[J].信息安全学报,已采用 [点击复制]
- zhanghui,zhangshengbing,luozhenjie,yanglibin,mudejun,majianfeng.A Review of the Current State of Research and Application Challenges in Federated Large Models[J].Journal of Cyber Security,Accept [点击复制]
|
|
| 摘要: |
| 联邦大模型通过融合联邦学习与大模型技术，可在保障数据隐私安全的前提下实现复杂任务的多方协同训练与知识推理，成为当前人工智能安全领域的研究热点。本文全面梳理了联邦大模型的研究现状，系统分析了其在实际应用中所面临的技术瓶颈问题，包括数据异质强、通信开销大、边缘算力弱以及安全风险升级等关键挑战，从现有策略设计、机制构建等角度进行了深入探讨，并针对学界流行的解决方案进行了系统性评述。其中，在数据异质性缓解方面，现有研究主要沿聚类分组、提示学习、参数高效微调等技术路径达到数据一致性增强目的；针对通信效率优化方面，现有研究重点集中于提示学习与模型分割等热点方向；在边缘算力受限条件下，现有解决方案主要聚焦于无反向传播、模型压缩、参数高效微调和动态资源调度等方法；在安全与隐私方向，本文从防御和攻击两个维度对相关研究进行了分析总结。进一步地，本文还从核心功能特点、关注重点、适用场景等角度，系统对比了现有联邦大模型开源框架异同；并给出了联邦大模型在法律、医疗、推荐系统等流行领域的应用案例，总结了当前联邦大模型尚待解决的共性问题及未来可能的研究趋势。通过对现有研究成果的深入分析，本文期望能为联邦大模型技术的优化和应用提供理论支持和实践指导。 |
| 关键词: 联邦大模型 数据异质性 通信开销 隐私保护 |
| DOI： |
| 投稿时间：2025-05-08修订日期：2025-08-06 |
| 基金项目:国家自然科学基金项目（面上项目，重点项目，重大项目） |
|
| A Review of the Current State of Research and Application Challenges in Federated Large Models |
|
zhanghui1, zhangshengbing1, luozhenjie2, yanglibin1, mudejun1, majianfeng3
|
| (1.School of Cybersecurity, Northwestern Polytechnical University;2.School of Automation, Northwestern Polytechnical University;3.School of Cyber Engineering, Xidian University) |
| Abstract: |
| By integrating federated learning with large model technology, federated large models enable multi-party cooperative training and knowledge sharing for complex tasks, all while ensuring robust data privacy and security. This powerful com-bination has emerged as a prominent research hotspot in the field of artificial intelligence security. This paper provides a comprehensive and in-depth review of the current state of research on federated large models and systematically analyzes the technical bottlenecks it faces in practical applications. These challenges include critical issues such as significant data het-erogeneity, high communication costs, limited edge computing power, and the escalation of security risks. In addition, this paper provides a detailed discussion of existing solutions from perspectives such as policy design and mechanism construction, and systematically reviews the popular solutions in the academic community. In terms of alleviating data heterogeneity, current research mainly focuses on techniques such as clustering, prompt learning, and parameter-efficient fi-ne-tuning to enhance data consistency. Regarding communication efficiency optimization, the existing research focuses on hot topics such as prompt learning and model segmentation. For scenarios with limited edge computing power, existing solutions mainly focus on methods such as non-backpropagation, model compression, parameter-efficient fine-tuning and dynamic resource scheduling. In the area of security and privacy, this paper analyzes and summarizes research from both defense and attack perspectives. Furthermore, the paper systematically compares existing open-source frameworks for federated large models, focusing on their core functionalities, areas of emphasis, and various application scenarios. It also presents application cases of federated large models in popular fields, such as law, healthcare, and recommendation systems. Moreover, the paper highlights common challenges faced by federated large models and summarizes possible future research directions. Through a thorough analysis of the existing body of research, this paper aims to provide both theoretical support and practical guidance for the optimization and application of federated large model technologies. |
| Key words: Federated Large Models Data Heterogeneity Communication Overhead Privacy Protection |
