Trans-GAN: A Deep Learning Paradigm for Multi-Type Anomaly Detection in Network Traffic
Muhammad Zeeshan, Maasooma · 2024
Identifying and mitigating network anomalies has become paramount with the exponential growth of IoT devices and networked computer systems. Intelligent and proactive systems are essential in today's digital landscape, yet existing detection methods struggle to cope with diverse and dynamic network behaviors. Consequently, proactive early detection has become increasingly complex, labor-intensive, and critical amidst escalating threats and intricate network protocol fields. To address these challenges, this study introduces a novel Trans-GAN deep learning model for multi-type anomaly detection. By combining generative adversarial networks with transformers, the Trans-GAN model learns the correlation between various network traffic behaviors and discovers interdependencies among parameters. We leverage the CICIoT23 dataset, which undergoes preprocessing, including normalization, labeling, and transformer-based encoding. To balance the dataset, we employ SMOTE-ENN and chunking mechanisms. The integration of a transformer self-attention mechanism with generative adversarial networks enables the Trans-GAN model to learn dynamic data patterns, adapt to changes, and detect multi-type attacks. The generator's loss function measures information loss between real and generated data, making it more suitable for creating synthetic tabular data. The feature tokenizer transforms categorical and numerical features into tokens, feeding them into the stacked transformer model. Each transformer layer analyzes features of individual data objects, enhancing the model's ability to understand and predict patterns. We validate the Trans-GAN model's effectiveness through extensive experimentation on the CICIoT23 dataset, achieving detection accuracies of 96.1% for DDoS attacks, 96.8% for DoS attacks, 94.8% for reconnaissance, 91.5% for brute-force attacks, and 90.3% for web-based attacks. Trans-GAN results significantly outperform other algorithms, confirming the effectiveness of the proposed Trans-GAN model for network traffic anomaly detection.