A Network Intrusion Detection Method Based on Parameter Optimization and Transformer

This paper proposes a high-performance network intrusion detection method based on parameter optimization and Transformer. To fully exploit the powerful sequence modeling capability of the Transformer architecture in network traffic analysis, we design a lightweight Transformer encoder with carefully tuned hyperparameters. The model projects each traffic flow record into a hidden space and performs multi-head self-attention to capture complex dependencies among features. Extensive experiments are conducted on the NSL-KDD datasets. By adopting AdamW optimizer, cosine annealing learning rate, and balanced sampling strategy, the proposed method achieves 99.0574% accuracy, 98.9862% precision, 99.0555% recall, and 99.0208% F1-score on the NSL-KDD dataset, significantly outperforming traditional machine learning and ordinary deep learning baselines. Comprehensive visualization results—including confusion matrix, probability distribution, t-SNE projection, and attention heatmap—fully demonstrate the superior classification performance and excellent interpretability of the model. The complete, reproducible code and all experimental figures are provided, offering an efficient and reliable solution for real-world network intrusion detection tasks.