Manet: Hybrid Mamba-Attention Model for Malicious Traffic Identification

The rapid evolution of sophisticated security evasion techniques and widespread adoption of encryption technologies have posed significant challenges to internet security, making effective malicious traffic identification crucial for network administrators. While researchers have proposed various approaches including machine learning and deep learning-based methods, these solutions face persistent challenges such as privacy leakage concerns, difficulties in acquiring training data, and insufficient model performance. To address these limitations, we present MAnet, a novel hybrid architecture integrating Mambastructured networks with attention mechanisms, designed to autonomously learn discriminative features of malicious network behaviors from raw traffic data. Comprehensive evaluations across six publicly available datasets demonstrate the model's superior performance: MAnet achieves top-performing results on five benchmark datasets and obtains comparable performance on the remaining dataset with only 0.15 % lower F1-score compared to the best existing approach. These experimental results not only validate the effectiveness of MAnet's architectural design but also suggest its strong potential for practical deployment in real-world network security applications. Furthermore, we conducted rigorous few-shot learning experiments to evaluate MAnet's adaptability. The results demonstrate comparable learning efficiency and generalization capabilities with best-performing counterparts. To the best of our knowledge, MAnet represents the first domainspecific pre-trained model explicitly designed and optimized for malicious traffic identification.