To overcome the limitations of traditional intrusion detection methods in dealing with high-dimensional sparse features, multiclass attack classification, and model robustness, this paper presents a fused multibranch intrusion detection model (FMB-IDM). The proposed framework combines three complementary deep learning components. First, an MLP branch equipped with a Squeeze-and-Excitation attention module is used to enhance feature representation by adaptively reweighting channel-wise information. Second, a TabTransformer-based branch is introduced to model complex dependencies among structured input features through self-attention. Third, a CNN-BiLSTM branch is employed to capture both local feature patterns and longer-range sequential relationships, where the CNN extracts local representations, and the two-layer bidirectional LSTM learns contextual dependencies. To make better use of the information learned by each branch, an attention-based fusion strategy is adopted to combine their outputs adaptively, which improves the model’s ability to distinguish different types of intrusion behaviors. Experiments conducted on the NSL-KDD, UNSW-NB15, and CIC-IDS2017 datasets demonstrate that the proposed model achieves strong and consistent performance in both binary and multiclass intrusion detection tasks. In particular, on CIC-IDS2017, FMB-IDM reaches an accuracy of 97.56% and a weighted F1-score of 0.98. In addition, the model maintains good inference efficiency, with an average latency of about 0.03 ms per sample under the experimental hardware configuration.
原文链接
分享
引用
