【Author】
Mugunthan, Vaikkunth; Rahman, Ravi; Kagal, Lalana
【Source】BLOCKCHAIN AND APPLICATIONS
【Abstract】Federated machine learning enables multiple clients to collectively train a machine learning model without sharing sensitive data. However, without proper accountability mechanisms, adversarial clients can weaken the collective model. BlockFLow is a fully decentralized, privacy-preserving, and accountable federated learning system. It introduces an Ethereum blockchain smart contract to coordinate a federated learning experiment and to hold clients accountable. BlockFLow rewards clients proportional to the quality of their individual contributions, does not reveal the underlying datasets, and is resilient to a minority of adversarial clients. Unlike existing systems, BlockFLow does not require a centralized test dataset, sharing of datasets between the clients, or any trusted entities. We evaluated BlockFLow on logistic regression models. Our results illustrate that BlockFLow successfully rewards honest clients and identifies adversarial clients. These results, along with blockchain costs that do not scale with model complexity, demonstrate the effectiveness of BlockFLow as an accountable federated learning system.
【Keywords】Blockchain accountability; Federated machine learning
【标题】BlockFLow：去中心化、隐私保护和负责任的联邦机器学习
【摘要】联邦机器学习使多个客户端能够共同训练机器学习模型，而无需共享敏感数据。然而，如果没有适当的问责机制，敌对客户可能会削弱集体模式。 BlockFLow 是一个完全去中心化、隐私保护和负责任的联邦学习系统。它引入了以太坊区块链智能合约来协调联邦学习实验并让客户承担责任。 BlockFLow 奖励与其个人贡献质量成比例的客户，不透露底层数据集，并且对少数对抗性客户具有弹性。与现有系统不同，BlockFlow 不需要集中的测试数据集、客户端之间的数据集共享或任何受信任的实体。我们在逻辑回归模型上评估了 BlockFlow。我们的结果表明，BlockFlow 成功地奖励了诚实的客户并识别了敌对的客户。这些结果，以及不随模型复杂性而扩展的区块链成本，证明了 BlockFLow 作为一个负责任的联邦学习系统的有效性。
评论