【Author】
Toyoda, Kentaroh; Mang, Allan N.
【Source】2019 IEEE INTERNATIONAL CONFERENCE ON BIG DATA (BIG DATA)
【Abstract】Recent technological evolution enables Artificial Intelligence (AI) model training by users' mobile devices, which accelerates decentralized big data analysis. In particular, Federated Learning (FL) is a key enabler to realize decentralized AI model update without user's privacy disclosure. However, since the behaviour of workers, who are assigned a training task, cannot be monitored, the state-of-the-art methods require a special hardware and/or cryptography to force the workers behave honestly, which hinders the realization. Furthermore, although blockchain-enabled FL has been proposed to give workers reward, any rigorous reward policy design has not been discussed. In this paper, to tackle these issues, we present a novel method using mechanism design, which is an economic approach to realize desired objectives under the situation that participants act rationally. The key idea is to introduce repeated competition for FL so that any rational worker follows the protocol and maximize their profits. With mechanism design, we propose a generic full-fledged protocol design for FL on a public blockchain. We also theoretically clarify incentive compatibility based on contest theory which is an auction-based game theory in economics.
【摘要】最近的技术发展使人工智能 (AI) 模型能够通过用户的移动设备进行训练，从而加速了分散的大数据分析。特别是，联邦学习（FL）是实现去中心化 AI 模型更新而无需用户隐私泄露的关键推动力。但是，由于无法监控分配给培训任务的工人的行为，因此最先进的方法需要特殊的硬件和/或密码学来强制工人诚实行事，这阻碍了实现。此外，尽管已经提出启用区块链的 FL 来给予工人奖励，但尚未讨论任何严格的奖励政策设计。在本文中，为了解决这些问题，我们提出了一种使用机制设计的新方法，这是一种在参与者理性行动的情况下实现预期目标的经济方法。关键思想是为 FL 引入重复竞争，以便任何理性的工人都遵循协议并最大化他们的利润。通过机制设计，我们为公共区块链上的 FL 提出了一个通用的成熟协议设计。我们还从理论上阐明了基于竞争理论的激励相容性，竞争理论是经济学中基于拍卖的博弈论。
评论