Secure Federated Learning for IoT using DRL-based Trust Mechanism

Federated learning (FL) has evolved to leverage a distributed dataset from numerous IoT devices to improve the performance of a Machine Learning (ML) model while preserving the privacy of device data. Client devices train a global model jointly and share local model updates with a central entity or a server. However, FL is vulnerable to a variety of adversarial attacks that target its security and privacy and lead to compromising the main FL task. In particular, devices can contribute unreliable local model updates due to poisoning attack, or unintentionally due to their limited resources. Therefore, identifying trustworthy and reliable devices to participate in FL task is a key security challenge. In this paper, we propose a reputation management mechanism based on Deep Reinforcement Learning (DRL) in order to optimize the selection and evaluation of reliable devices and improve the accuracy of the FL model. The experimental results show that the proposed DRL-based reputation management scheme can enhance the FL accuracy by 20% while requiring fewer training iterations when compared to conventional reputation-based methods.