Adversarially robust federated deep learning models for intrusion detection in IoT

Abstract

Ensuring the robustness, security, and privacy of machine learning is a pivotal objective, crucial for unlocking the complete potential of the internet of things (IoT). Deep neural networks have proven to be vulnerable to adversarial perturbations imperceptible to humans. These perturbations can give rise to adversarial attacks, leading to erroneous predictions by deep neural networks, particularly in intrusion detection within the IoT environment. This paper introduces a federated adversarial learning framework designed to protect both data privacy and deep neural network models. This framework consists of federated learning for data privacy and adversarial training on IoT devices to enhance model robustness. The experiments show that adversarial training at the Fog node devices significantly improves the robustness of a federated learning model against adversarial attacks when compared to normal training. Furthermore, the proposed adversarial deep federated learning model is validated using the Edge-IIoTset dataset, achieving an accuracy rate of 91.23% in the detection of attacks.