TD-MDB: A Truth-Discovery-Based Multidimensional Bidding Strategy for Federated Learning in Industrial IoT Systems

As Industrial Internet of Things (IIoT) systems have superior distributed characteristics, federated learning (FL) has gradually become a mainstream distributed computing paradigm in recent years, promoting its intelligent development. A key factor of FL is the model accuracy, which relies on the truth of the participating industrial edge device (IED)'s parameter. However, some untrusted IEDs can contaminate models by fabricating false FL parameters, thereby deceiving rewards, known as model poison attack (MPA). In this article, a truth-discovery-based multidimensional bidding (TD-MDB) strategy is proposed to obtain truthful parameters at a low cost for FL. In the TD-MDB strategy, parameter server (PS) will randomly assign the lightweight trust evaluation model to test the trust of each IED according to the accuracy of IED's calculation. Then, in order to resist MPA and inference attack, we use an improved clustering algorithm to divide the trusted IEDs into different clusters based on their different trust and stability. Finally, a device selection strategy is proposed to select cost-effective IEDs for model training based on Asynchronous Advantage Actor-Critic (A3C) algorithm. In the TD-MDB strategy, the IEDs that provide more resources, high trust, and low bid will be selected to train the model and maximize the real revenue of the system with high accuracy. We finally carry out extensive evaluations, where results demonstrate that the TD-MDB strategy can effectively provide an efficient offloading strategy, resist MPA and ensure that truth computing for FL, which is better than the state-of-the-art strategies in model accuracy, system revenue, and other performance indicators.