Lightweight Diffusion Model for Synthesizing Malicious Network Traffic
Fuhao Li, Hongyu Wu, Jielun Zhang
Abstract
Malicious traffic identification is essential for main-taining network security by detecting and preventing intrusive activities. With the emergence of edge computing applications, identifying malicious traffic on edge nodes has become critical. However, the scarcity of high-quality malicious traffic data poses a significant challenge in training effective AI models for this purpose. Additionally, the low computational power of edge nodes hinders the use of generative models, such as diffusion models, for data augmentation to facilitate model training. This paper addresses these issues by proposing a novel lightweight diffusion model design specifically tailored for synthesizing ma-licious traffic data to be used as training samples, enabling robust AI models. Our approach leverages depth wise separable convolutions to reduce computational overhead while maintaining the quality and diversity of the generated data. We evaluated our approach using the USTC- TFC2016 dataset, demonstrating its ability to produce realistic and diverse malicious traffic data with significantly reduced computational requirements.