Botnet Traffic Detection via Deep Neural Networks with Multi-Category Feature Fusion and Multi-Head Self-Attention Mechanism

With the continuous evolution of botnet attack techniques, traditional detection methods based on signature matching and single-feature analysis are increasingly inadequate in handling complex and dynamic network threats, resulting in limited accuracy and robustness. Effectively integrating heterogeneous features from multiple sources to achieve high-precision botnet traffic detection has become a critical challenge in the field of network security. To address this issue, this paper proposes a multimodal deep neural network architecture that fuses statistical and categorical features. By incorporating embedding layers and a multi-head self-attention mechanism, the model enables collaborative modeling and deep interaction learning of heterogeneous features. Experimental results on the CTU-13 and Bot-IoT datasets demonstrate that the proposed model achieves an accuracy of $\mathbf{9 9. 4 8 \%}$ and an F1-score of $\mathbf{9 8. 9 2 \%}$, significantly outperforming traditional machine learning methods and common deep learning baselines. The proposed model substantially improves the accuracy and generalization of botnet detection, providing a robust technical foundation for security defense in complex network environments. It holds significant practical value and broad prospects for application.