Stable Discrete Segmented Reverse Diffusion Model for Solving Class Imbalance in Malicious Websites Detection

In order to address the class imbalance issue experienced during training for malicious website detection, we developed a deep generative model based on the diffusion model that was able to generate the features of malicious websites. The use of the binary encoder method effectively solved the problem experienced with traditional diffusion models, which are unable to generate the discrete and integer features of malicious websites. Moreover, we modified the reverse diffusion algorithm so that, in the final step of the reverse diffusion process, the original data was restored by predicting the noise rather than by sampling the original data distribution through calculating mean and variance. This method enhanced the stability of the features generated and improved their quality.