A DISTRIBUTED LEDGER-ENABLED COLLABORATIVE INTELLIGENCE ARCHITECTURE INCORPORATING DUAL CONFIDENTIALITY PRESERVATION AND TRUST-WEIGHTED AGREEMENT

Abstract

In IIoT situations, federated learning (FL) is a way to use industrial data that protects privacy. At the same time, adding blockchain to federated learning training makes it more trustworthy. But there are still some big problems with current blockchain-based FL frameworks: 1) The current consensus mechanisms don't do a good job of filtering out bad devices, which lets low-quality participants mess with global model training and make the model less robust; 2) Current privacy budget strategies are too simple, making it hard to find a balance between protecting privacy during statistical queries and gradient updates. Strong privacy protection lowers model accuracy, while weak protection doesn't protect against poisoning attacks. This paper proposes ShieldDFL, a blockchain-based federated learning framework with dual privacy protection and reputation-driven consensus, to solve these problems. This method uses a hybrid consensus mechanism based on LSTM-based reputation scoring to dynamically assess both short-term and long-term device contributions. This makes it possible to choose the best devices with accuracy. At the same time, it adds a new dual privacy budget mechanism that uses differential privacy for both statistical queries and gradient updates. This keeps privacy strong while keeping the model's performance high. The proposed method lowers the chances of bad devices getting into the consensus pool to 1.5%, lowers the success rates of SAR and BASR attacks to 5.8% and 2.1%, respectively, and keeps the model's accuracy high at 98.1% on MNIST and 87.6% on CIFAR-10. In general, the proposed framework does a good job of getting around the security and privacy problems that come with blockchain-based federated learning. It offers a fast and flexible way for decentralised and trustworthy collaboration in IIoT situations.