Learning to Predict Tendency of CVE Vulnerability

Predicting the trend of software vulnerabilities enables proactive warnings, aiding in mitigating security incident losses. Capturing the long-term dependencies among data is crucial for achieving highly accurate predictions. Leveraging the remarkable data relationship-capturing capabilities of learning-based methods demonstrated in various domains, we implemented several learning-based approaches to forecast the outbreak trend of software vulnerabilities. The results show that Informer, Autoformer, and DLinear achieved accuracy rates exceeding 90%. Informer, with its residual connection module, obtained significant prediction results with fewer data points. Autoformer, benefiting from its data decomposition structure, performed better in datasets with periodic patterns. DLinear, on the other hand, exhibited superior performance in relatively stable datasets. However, LSTM performed poorly across multiple datasets due to the cumulative error effect.