Adversarial Attack Detection in Wireless Networks Using Deep Learning Based Capsule Networks

Abstract

The vulnerability of wireless networks to misclassification and security oversight is increased by adversarial attacks using Deep Learning (DL)-based Intrusion Detection Systems (IDS). Because they are unable to distinguish small fluctuations in signal data and retain spatial hierarchy, traditional neural networks prefer Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), which are highly prone to adversarial perturbations. In this research, a Capsule Network (CapsNet)-based framework detecting adversarial attacks in wireless networks is proposed. The network uses RF-aware capsule representations and dynamic routing to improve hierarchical feature learning and robustness against adversarial perturbations. To successfully search for adversarial distortion, the proposed model extracts significant network traffic features, encodes spatial hierarchy via primary and digit capsules, and uses a reconstruction loss function. To improve the model's robustness against sophisticated attack tactics, an exploratory quantum-assisted CapsNet implementation for preliminary investigation is also included. Experimental evaluation on benchmark wireless intrusion datasets shows that CapsNets outperform conventional CNNs and Long Short-Term Memory (LSTM) models, achieving 98.3% accuracy under normal conditions and maintaining 91.3% precision even under adversarial attack, compared to 87.5% and 72.4% for CNNs and 90.2% and 78.6% for LSTMs, respectively. In addition, compared with traditional DL models, CapsNets show a 34% improvement in robustness metrics. The proposed CapsNet system achieved 98.3% detection accuracy and a reduced False Positive Rate across several adversarial attack scenarios.