Abstract:

To address the limitations of current background traffic generation methods in power communication—particularly in modeling protocol behaviors, capturing temporal dependencies, and controlling traffic category distributions—this paper proposes a background traffic generation approach based on diffusion models and bidirectional flow (DMBF). By employing transforms basic flow data into an intuitive picture (FlowPic), we extract bidirectional session images featuring directionality, temporality, and packet-length coupling characteristics. This is combined with a Transformer for temporal modeling. A conditional control mechanism is introduced to adjust the generation ratios of different traffic types, enabling the diffusion model to generate background flows under guided conditions. To evaluate the practicality and generalizability of the proposed method, experiments are conducted on datasets comprising both publicly available traffic samples and real-world network communication data, covering a range of typical business scenarios and interaction patterns. Experimental results show that DMBF outperforms traditional generative adversarial network approaches in terms of generation accuracy and distributional consistency. JSD decreased to 28.89%, with MAE and RMSE at 26.24% and 30.91%, respectively.

Key words: power communication, cyber security, traffic generation, diffusion model, feature extraction, deep learning