A Transformer-Based Intrusion Detection Approach with Fast Gradient Sign Method Adversarial Training

Intrusion detection systems are crucial for maintaining the security of network infrastructures, yet they remain vulnerable to sophisticated and adversarial attacks.To address this issue, we propose a novel intrusion detection method that combines the transformer architecture with Fast Gradient Sign Method (FGSM) adversarial training.The transformer-based network is specifically designed to process network traffic data, utilizing a transformer encoder with multi-head self-attention mechanisms and position-wise feed-forward layers.The core innovation lies in replacing standard training procedures with FGSM adversarial training, where the model is trained on both clean and adversarial examples.This enhances the system's ability to detect and resist adversarial attacks.The network's performance is evaluated under varying values of the adversarial rate parameter (lambda), with the best results achieved at lambda = 0.5.The model demonstrates a high classification accuracy of 99.7321% on the training data (using the NSL-KDD dataset), 99.603% on the testing data, and 99.4641% on adversarial data.These findings demonstrate the method's robustness and reliability, providing an effective solution for secure and efficient intrusion detection.