Hybrid Anomaly Detection Using Vision Transformers and Convolutional Neural Networks

This paper proposes a hybrid anomaly detection model that integrates Vision Transformers(ViTs) and Convolutional Neural Networks(CNNs) for graph-based classification of network behaviors. ViTs capture global dependencies through self-attention, whereas CNNs extract localized patterns via convolutional kernels. By combining these complementary features, the model improves detection accuracy for tasks such as intrusion detection and phishing website identification. Evaluations on four benchmark datasets-CIC-IDS2017, Kaggle Phishing, UNSW-NB15, and CICIoT2023-yield accuracies of 99.92%, 99.91%, 100.0%, and 99.99%, respectively. Rather than relying on aggregated inputs, the proposed method leverages independent instance classification, thereby providing the flexibility required for real-time anomaly detection across diverse domains.