A Comprehensive Study on Lightweight Convolution Techniques for Malicious Traffic Synthesis in Diffusion Models

Network security relies on effective and accurate malicious traffic detection, which is increasingly important for edge devices. As computing resources are distributed across many devices, detecting malicious traffic at the edge becomes crucial. Additionally, the lack of high-quality malicious traffic data hinders the effectiveness of AI models in training. Furthermore, the limited computational power of edge devices requires lightweight generative models, making it challenging to implement diffusion models for data augmentation. In this paper, we conduct a comprehensive study on lightweight convolution techniques for malicious traffic synthesis in diffusion models. We evaluate the performance of three classic lightweight convolutional architectures in generating malicious traffic data. Two popular datasets, USTC-TFC2016 and N-BaIoT, are utilized to train our generative models, with the generated data being employed to train various classifiers for a thorough performance evaluation. The experimental results demonstrate that the investigated lightweight approaches can efficiently generate high-quality and diverse malicious traffic data for enhancing the performance of intrusion detection systems while the computational complexity is reduced to around 19.69% of the baseline.