Enhancing Cyber Threat Detection in Tweets Using Diffusion Models and Convolutional Neural Networks

This study looks at how two advanced methods, DDPM-v4 and uViT, work with CNN models to find cyber threats from X data. Two datasets were used—one dataset taken from GitHub and another dataset created using the X Developer API. The dataset created had two sizes: small with 500 rows and larger with 2500 rows. The study compares models with diffusion techniques (DDPM-v4 and uViT) and without diffusion techniques. The results show that DDPM-v4 performed the best overall, with the highest accuracy and F1 score. For example, on the X API dataset with 2500 rows, DDPM achieved 98.20% accuracy. The uViT model also worked well but had slightly lower results. However, when the models were tested on small datasets of 500 rows, the accuracy was 99.9% to 100%, which is not realistic. It showed that small datasets cause overfitting, where the model memorizes data instead of learning. Larger datasets, like 2500 rows, gave more realistic and trustworthy results. The study proves that using advanced diffusion methods like DDPM-v4 and uViT can improve the model's ability to detect cyber threats. It also shows that bigger datasets are very important for getting real and useful results. In the future, more large-scale data can help make these methods even better for real-world applications.