Abstract-In today’s era, Cyber Physical Systems (CPS) with Internet of Things (IoT) have gained immense importance as integral building blocks of smart healthcare systems, automation industries, intelligent transport systems, and smart city architectures. High connectivity and distribution architecture make such CPSs susceptible to highly complex adversarial attacks which can manipulate sensor inputs, network flows, and machine learning-based predictions. The traditional intrusion detection methods give primary focus on classification accuracy while being susceptible to adversarial attacks. This causes significant degradation in performance of such detectors leading to poor system reliability. In order to overcome such challenges, this paper introduces a novel deep learning based method known as Hybrid Contrastive Adversarial Defense Network (HCAD-Net). The proposed HCAD-Net is a new intrusion detector which utilizes supervised contrastive learning of representation, adversarial alignment of features, adaptive attention encoding of contextual IoT traffic characteristics, and uncertainty-based classification. Firstly, adversarial augmentation is applied on normal samples to generate perturbed samples to simulate real-world attacks. Then, dual view contrastive encoder is used to learn robust feature representations through minimizing intra-class distances while maximizing inter-class separability. The performance of the proposed HCAD-Net model has been evaluated through an experiment on the CICIoT2023 benchmark dataset, which shows an accuracy of 99.21%, precision of 98.94%, recall of 98.81%, F1-score of 98.87%, and an attack detection rate of 99.05%. These results show that the model has performed better than CNN-based, LSTM-based, and transformer-based intrusion detection techniques.
 

Deep Contrastive Learning, Adversarial Defense, Cyber-Physical Systems, IoT Security, Intrusion Detection