ACS: Accuracy-based client selection mechanism for federated industrial IoT

This study proposes secure federated learning (FL)-based architecture for the industrial internet of things (IIoT) with a novel client selection mechanism to enhance the learning performance. In order to secure the FL architecture and ensure that available clients are trustworthy, a certificate authority (CA) is adopted. In traditional FL, an aggregation technique known as federated averaging (FedAvg) is utilized to collect local model parameters by selecting a random subset of clients for the training process. However, the random selection may lead to uncertainty and negatively influence the overall FL performance. Moreover, state-of-the-art studies on client selection mainly rely on client’s additional information, which raises a privacy issue. Therefore, a novel client selection mechanism based on client evaluation accuracy called ACS is introduced in this work to improve FL performance while preserving client privacy. Unlike other client selection methods, ACS relies only on the updated local parameter, which is evaluated in the FL server. The proposed ACS considers the highest-performing clients to fasten the convergence time in the FL. Based on the extensive performance evaluation performed in this work using MNIST and F-MNIST datasets with non-independent identically distributed (non-IID) conditions, the adoption of ACS successfully improved the overall performance of FL in terms of accuracy and F1-score with an average of 4.62%. Furthermore, comparative analysis shows that the proposed ACS can achieve specific accuracy with 2.29% lower communication rounds and stable performance compared to other client selection mechanisms.