An Innovative Framework for Adversarial Attack Detection Using AI in Federated Learning

Abstract

The Internet of Things (IoT) consists of millions of physical objects connected each other through internet connection and performs tasks independently with less human interference. Although IoT has numerous benefits, it is prone to malicious attacks. In recent years, the evolution of advanced technologies such as machine learning (ML) and deep learning has emerged as a promising solution for attack detection. In this study, a novel attack detection framework was proposed by combining Bobcat optimized Autoencoder-based Capsule Network (BoAEbC) with the Federated Learning (BoAEbC-FL) to detect adversarial attacks in the IoT environment. The proposed methodology operates in a decentralized manner on data for detecting attacks, which boosts network’s security and data privacy. The presented study utilized BoAEbC as the base model for attack detection and the central FL server aggregates the model’s weights from multiple local models using federated average (FedAvg) algorithm, providing collaborative learning and ensures data privacy. The BOA utilized in the designed algorithm fine-tunes the constraints of the autoencoder-based capsule network, which enhances its learning capabilities and adaptability. The presented strategy was trained and validated using the public IoT database in Python platform and the experimental results are determined in terms of metrics like accuracy, specificity, f-measure, sensitivity, precision, etc. The implementation results are validated with the conventional attack detection models through comparative assessment and it highlighted that the designed method offered superior performances than others.