信息网络安全 ›› 2024, Vol. 24 ›› Issue (1): 1-13.doi: 10.3969/j.issn.1671-1122.2024.01.001

• 区块链与可信交易 • 上一篇    下一篇

基于零知识证明和区块链的联邦学习激励方案

吴昊天1, 李一凡1, 崔鸿雁2, 董琳3()   

  1. 1.华南理工大学计算机科学与工程学院,广州 510006
    2.郑州信大先进技术研究院,郑州 450001
    3.国家计算机网络与信息安全管理中心,北京 100029
  • 收稿日期:2023-08-26 出版日期:2024-01-10 发布日期:2024-01-24
  • 通讯作者: 董琳 E-mail:donglin@cert.org.cn
  • 作者简介:吴昊天(1980—),男,江苏,副教授,博士,CCF会员,主要研究方向为可逆信息隐藏、隐私保护、口令猜解和区块链|李一凡(1984—),男,湖南,硕士研究生,主要研究方向为区块链、零知识证明|崔鸿雁(1978—)男,陕西,讲师,硕士,主要方向为信息安全、大数据应用|董琳(1985—),女,山西,高级工程师,博士,主要研究方向为互联网金融安全、网络信息安全
  • 基金资助:
    国家重点研发计划(2022YFB2702503);广东省自然科学基金(2021A1515011798);河南省网络空间态势感知重点实验室开放课题(HNTS2022017)

Federated Learning Incentive Scheme Based on Zero-Knowledge Proofs and Blockchain

WU Haotian1, LI Yifan1, CUI Hongyan2, DONG Lin3()   

  1. 1. School of Computer Science and Engineering, South China University of Technology, Guangzhou 510006, China
    2. Zhengzhou Xinda Institute of Advanced Technology, Zhengzhou 450001, China
    3. National Computer Network Emergency Response Technical Team/Coordination Center of China, Beijing 100029, China
  • Received:2023-08-26 Online:2024-01-10 Published:2024-01-24
  • Contact: DONG Lin E-mail:donglin@cert.org.cn

摘要:

在跨孤岛联邦学习中,各参与者对最终训练出的模型贡献各异,如何评估他们的贡献并给予适当激励,成为联邦学习研究中一项关键问题。目前的激励方法主要着眼于奖励有效模型更新的参与者,同时惩罚不诚实者,侧重于激励计算行为。然而,参与者所提供的数据质量同样影响学习效果,但现有方法未充分考虑数据质量,并缺乏鉴定数据真实性的手段。为提升激励的准确性,需对参与者数据质量进行评估。通过融合零知识证明与区块链技术,文章提出一种评估参与者数据质量的协议,构建了全新联邦学习激励方案。该方案可在不泄露明文数据的前提下,评估参与者所用数据集质量,通过区块链系统向合格参与者发放激励,拒绝不合格者。实验证实,在部分用户提供虚假数据的情况下,该方案仍能准确给出激励结果,同时提升联邦学习模型的准确率。

关键词: 零知识证明, 区块链, 激励机制, 联邦学习, 数据质量评估

Abstract:

In cross-silo federated learning, participants contribute differently to the final trained model. Evaluating their contributions and providing appropriate incentives has become a key issue in federated learning research. Current incentive methods primarily focus on rewarding participants who provide valid model updates while penalizing dishonest ones, emphasizing incentivizing computational behavior. However, the quality of data provided by participants also affects learning outcomes, yet existing methods inadequately consider data quality and lack means to verify data authenticity. To enhance incentive accuracy, it is necessary to evaluate the quality of participants' data. This paper introduced, for the first time, a protocol for assessing the quality of participants' data by integrating zero-knowledge proofs and blockchain technology, leading to a novel federated learning incentive scheme. This scheme can assess the quality of participants' datasets without disclosing plaintext data, utilizing blockchain systems to provide incentives to eligible participants while excluding those who don't meet the criteria. Experimental results confirm that even in scenarios where some users provide falsified data, this scheme remains capable of delivering accurate incentive results, while simultaneously improving the accuracy of the federated learning model.

Key words: zero-knowledge proofs, blockchain, incentive mechanism, federated learning, data quality assessment

中图分类号: