A Trustworthy Federated Learning Model: Client Selection for IoT Edge Networks
Merve Telçeken, Elif Bozkaya · 2024
The interaction between the Internet of Things (IoT) and edge computing plays a critical role in processing and analyzing massive amounts of data. However, due to the malicious attacks of underlying IoT edge networks, the collected data can be untrustworthy, and designing a robust and trustworthy attack detection scheme for cloud-edge collaborative architecture becomes a challenge. In addition, malicious attacks also threaten data privacy. To address the data privacy concern and the need for cooperation between clients, federated learning is a cutting-edge machine learning solution, where local models are trained on clients and the global model is jointly constructed with local parameters. In this regard, we propose a trustworthy client selection scheme for IoT edge networks. Specifically, we first devise a trust management mechanism and calculate the trust scores of clients by using the Isolation Forest method. Then, a trust-based federated learning algorithm is proposed to protect data privacy. In this algorithm, trustworthy clients are selected to constitute the parameters of the global model. Extensive experiments are conducted to validate the trustworthy client selection model against malicious attacks. As a case study, Distributed Denial of Service (DDoS) attack is analyzed and system performance is compared with traditional benchmark methods in terms of accuracy, recall, precision, and F1 score.