Leveraging Large Language Models For Automated Software Vulnerability Detection and Analysis

Abstract

The growing integration of software across various sectors has amplified the urgency to develop effective mechanisms for safeguarding systems against sophisticated cyberattacks. This thesis investigates the application of modern artificial intelligence specifically transformer-based models for improving software vulnerability detection. The research focuses on designing intelligent algorithms capable of autonomously analyzing source code and identifying subtle security flaws. It emphasizes automated detection and fine-grained classification of vulnerabilities. Our methodology includes constructing high-quality labeled datasets and training advanced models such as BERT and RoBERTa to extract threat patterns and detect weaknesses at the line level and across multiple vulnerability categories. Experimental results show that transformer-based models significantly outperform traditional methods in both detection accuracy and analysis efficiency. Moreover, the practical deployment of our model in real-world scenarios confirms its utility in supporting software security analysts. This work thus provides a meaningful contribution to the field of cybersecurity, offering a scalable and robust solution to the persistent challenges in securing modern software systems.