A Novel and Efficient Multi-scale Spatio-Temporal Residual Network for Multi-class Intrusion Detection

Nan Li, Zhaojian Gao, Jiabin Ye, Wei Tang, Xun Che, Yadang Chen

Abstract: With the development of network devices, the network traffic presents high-dimensional, enormous as well as complex characteristics, and the network threats and attacks continue to intensify. Existing network intrusion detection models tend to disregard the extraction and learning of the temporal features of data, which will greatly affect the accuracy of network intrusion detection models. To address the shortcomings of the single structure and the inability to comprehensively learn features, we propose a novel Transformer-based network intrusion detection method, which integrates the degree of importance of traffic features and the fact that intrusion detection data has temporal and spatial characteristics. Specifically, firstly, we adopt the Transformer to perform feature extraction and construct global correlations on the input data, after that, we utilize the improved Inception to extract multi-scale features and weight the spatial features at different scales using self-attention module, in addition to that, BiGRU is employed to enhance the temporal features. Finally, the proposed model is validated on the publicly available CIC-IDS-2017 and CIC-DDoS-2019 datasets, which verifies that the proposed method outperforms the existing state-of-the-art models in terms of performance by four evaluation metrics, and also shows significant performance improvement in binary as well as multi-class classification tasks compared to other state-of-the-art methods, which proves the efficiency and effectiveness of our method.