基于TabTransformer的网络入侵检测系统研究与实现

Research and Implementation of a Network Intrusion Detection System Based on TabTransformer

摘要: 随着网络攻击手段日益复杂多样，传统入侵检测方法在复杂流量环境下面临性能瓶颈。针对网络流量数据中高维特征及复杂特征交互问题，文章设计并实现了一种基于TabTransformer的端到端网络入侵检测系统，通过引入特征嵌入和自注意力机制，捕获网络流量数据中复杂的非线性关系。基于CIC-IDS2017和UNSW-NB15公开数据集的对比实验表明，该系统在准确率、F1-Score、PR-AUC等指标上均优于CNN、RNN、LSTM等典型深度学习模型，在两个数据集的F1-Score均达到0.98以上，可为Transformer在入侵检测中的应用提供了一定参考。

Abstract: As the methods of network attacks become increasingly complex and diverse, traditional intrusion detection methods are limited in performance in complex traffic environments. To address the issues of high-dimensional features and complex feature interactions in network traffic data, this paper designs and implements an end-to-end network intrusion detection system based on TabTransformer. By introducing feature embedding and self-attention mechanisms, it captures the complex nonlinear relationships in network traffic data. Comparative experiments based on the public datasets CIC-IDS2017 and UNSW-NB15 show that this system outperforms typical deep learning models such as CNN, RNN, and LSTM in terms of accuracy, F1-Score, PR-AUC, etc. The F1-Score of this system in both datasets reaches above 0.98, providing certain references for the application of the Transformer in intrusion detection.