FastEmbed: Predicting vulnerability exploitation possibility based on ensemble machine learning algorithm.

Area of Science:

• Cybersecurity
• Machine Learning
• Vulnerability Management

Background:

• Increasing number of disclosed vulnerabilities necessitates efficient prioritization.
• Existing methods struggle with diverse threat intelligence and context in vulnerability data.
• Need for advanced models to predict exploitability and focus on critical vulnerabilities.

Purpose of the Study:

• To develop a more general and effective exploit prediction model.
• To improve the prediction of exploited vulnerabilities using advanced text embedding techniques.
• To enhance cybersecurity by enabling quicker exclusion of non-exploitable vulnerabilities.

Main Methods:

• Proposed fastEmbed model combining fastText and LightGBM algorithms.
• Utilized vulnerability-related text embeddings to capture contextual meaning.
• Trained and evaluated on extremely imbalanced datasets, outperforming baseline models.

Main Results:

• fastEmbed demonstrated superior generalization and prediction abilities, outperforming benchmarks by 6.283%.
• Achieved an F1 score of 0.586 for predicting exploits in the wild, a 33.577% improvement.
• Effectively learned embeddings from vulnerability text, improving exploitability description.

Conclusions:

• The fastEmbed model offers a significant advancement in exploit prediction.
• It enhances the ability to identify and prioritize exploitable vulnerabilities.
• The model provides a more effective approach to cybersecurity defense and resource allocation.