A Novel Autoencoder based Federated Deep Transfer Learning and Weighted k-Subspace Network clustering for Intelligent Intrusion Detection for the Internet of Things

Abstract

Federated Learning (FL) has established as a potentially effective practice for cyberattack identification in the last decade, particularly for Internet-of-Things (IoT) structures. FL can increase learning effectiveness, lower transmission overheads, and enhance intrusion detection system (IDS) privacy by spreading the learning process amongst IoT gateways. The absence of labeled data and the distinction of data features for training pose significant obstacles to the deployment of FL in IoT networks. In this research, suggest an Autoencoder based Deep Federated Transfer Learning (ADFTL) to conquer these obstacles. Specifically, Create an ADFTL model utilizing two AutoEncoders (AEs) as the basis. Initially the supervised mode is employed to train the first AE (AE1) on the source datasets while the unsupervised mode is employed to train the second AE (AE2) on the target datasets without label information. The bottleneck layer, or latent representation, of AE2 is forced via the transfer learning method in an effort to resemble the latent representation of AE1. Subsequently, assaults in the input in the target domain are identified employing the latent representation of AE2. Particularly, Weighted k-Subspace Network (WkSNC) clustering is proposed for clustering the dataset and Boosted Sine Cos method (BSCM) is used for feature selection. The requirement that the network datasets utilized in current studies have identical properties is significant since it restricts the effectiveness, adaptability, and scalability of IDS. Nonetheless, the suggested structure can tackle these issues by sharing the "knowledge" of learning among distinct deep learning (DL) simulations, even in cases when their datasets possess dissimilar features. Comprehensive tests on current BoT-IoT datasets demonstrate that the suggested structure can outperform the most advanced DL-based methods by more than 6 %