Cross-device federated learning (CDFL) systems enable fully decentralized training networks whereby each participating device can act as a model-owner and a model-producer. CDFL systems need to ensure fairness, trustworthiness, and high-quality model availability across all the participants in the underlying training networks. This article presents a blockchain-based framework, TrustFed, for CDFL systems to detect the model poisoning attacks, enable fair training settings, and maintain the participating devices’ reputation. TrustFed provides fairness by detecting and removing the attackers from the training distributions. It uses blockchain smart contracts to maintain participating devices’ reputations to compel the participants in bringing active and honest model contributions. We implemented the TrustFed using a Python-simulated federated learning framework, blockchain smart contracts, and statistical outlier detection techniques. We tested it over the large-scale industrial Internet of things dataset and multiple attack models. We found that TrustFed produces better results regarding multiple aspects compared with the conventional baseline approaches.

中文翻译：

---

TrustFed：IIoT 中公平可信的跨设备联合学习框架

跨设备联合学习 (CDFL) 系统支持完全分散的训练网络，其中每个参与的设备都可以充当模型所有者和模型生产者。CDFL 系统需要确保底层训练网络中所有参与者的公平性、可信性和高质量模型可用性。本文介绍了一个基于区块链的框架 TrustFed，用于 CDFL 系统检测模型中毒攻击、启用公平的训练设置并维护参与设备的声誉。TrustFed 通过从训练分布中检测和移除攻击者来提供公平性。它使用区块链智能合约来维护参与设备的声誉，以迫使参与者带来积极和诚实的模型贡献。我们使用 Python 模拟的联合学习框架、区块链智能合约和统计异常值检测技术实现了 TrustFed。我们在大规模工业物联网数据集和多种攻击模型上对其进行了测试。我们发现，与传统的基线方法相比，TrustFed 在多个方面产生了更好的结果。