.png){kind=icon}
.png){kind=icon}
Return
Diffusion model-based network packet synthesis using inter-packet difference learning
.png){kind=icon}
PRE
AI
DOI：10.23919/comex.2025XBL0132.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In recent years, network analysis has increasingly used packet data analysis for machine learning. However, packet data accumulation has issues, including difficulties in data accumulation and insufficient diversity in attack patterns. To address these issues, researchers have increasingly turned to synthetic data generation methods as an alternative to data accumulation approaches for training machine learning models. However, conventional GAN-based synthetic data generation methods have limitations, particularly statistical characteristic inconsistencies due to training instability and lack of correlations both within packet fields and between consecutive packets. This paper proposes an approach that transforms packet data into structured image representations and generates differential image data using a conditional diffusion model on the basis of previous packet data. The proposed differential representation method excludes unchanged fields from learning and focuses specifically on the varying components that capture inter-packet relationships. Evaluation experiments conducted on the CICIDS 2017 dataset demonstrate the proposed approach improves over conventional methods in both statistical distribution similarity metrics and classification difficulty assessments.
Keywords:
network packet generation
diffusion model
synthetic data