【Author】
Bao, Xianglin; Su, Cheng; Xiong, Yan; Huang, Wenchao; Hu, Yifei
【Source】5TH INTERNATIONAL CONFERENCE ON BIG DATA COMPUTING AND COMMUNICATIONS (BIGCOM 2019)
【Abstract】Federated learning (shorted as FL) recently proposed by Google is a privacy-preserving method to integrate distributed data trainers. FL is extremely useful due to its ensuring privacy, lower latency, less power consumption and smarter models, but it could fail if multiple trainers abort training or send malformed messages to its partners. Such mis-behavior are not auditable and parameter server may compute incorrectly due to single point failure. Furthermore, FL has no incentive to attract sufficient distributed training data and computation power. In this paper, we propose FLChain to build a decentralized, public auditable and healthy FL ecosystem with trust and incentive. FLChain replace traditional FL parameter server whose computation result must be consensual on-chain. Our work is not trivial when it is vital and hard to provide enough incentive and deterrence to distributed trainers. We achieve model commercialization by providing a healthy marketplace for collaborative-training models. Honest trainer can gain fairly partitioned profit from well-trained model according to its contribution and the malicious can be timely detected and heavily punished. To reduce the time cost of misbehavior detecting and model query, we design DDCBF for accelerating the query of blockchain-documented information. Finally, we implement a prototype of our work and measure the cost of various operations.
【Keywords】blockchain; federated learning; incentive; decentralize; trust
【标题】FLChain：具有信任和激励的可审计联邦学习区块链
【摘要】Google 最近提出的联邦学习（简称 FL）是一种集成分布式数据训练器的隐私保护方法。 FL 非常有用，因为它可以确保隐私、更低的延迟、更少的功耗和更智能的模型，但如果多个培训师中止培训或向其合作伙伴发送格式错误的消息，它可能会失败。这种不当行为是不可审计的，并且参数服务器可能由于单点故障而计算不正确。此外，FL 没有动力去吸引足够的分布式训练数据和计算能力。在本文中，我们提出 FLChain 以建立一个去中心化、可公开审计、健康的 FL 生态系统，具有信任和激励。 FLChain 取代了传统的 FL 参数服务器，其计算结果必须在链上达成共识。当为分布式培训师提供足够的激励和威慑至关重要且难以做到时，我们的工作并非微不足道。我们通过为协作训练模型提供健康的市场来实现模型商业化。诚实的训练者可以根据其贡献从训练好的模型中获得公平分配的利润，恶意的可以被及时发现并严厉惩罚。为了降低不当行为检测和模型查询的时间成本，我们设计了 DDCBF 来加速区块链文档信息的查询。最后，我们实现了我们工作的原型并测量了各种操作的成本。
评论