On-Premise CodeBERT-Driven Model for Vulnerability Detection in Source Code

Abstract

Security vulnerabilities in software systems remain a major concern in modern computing, especially as applications grow in complexity and are increasingly integrated into critical infrastructures. Traditional vulnerability detection methods such as static code analysis tools and manual inspection face limitations in accuracy, scalability, and privacy preservation, since they often rely on cloud-based solutions that expose sensitive source code. This paper presents an on premise Artificial Intelligence (AI)-based model for vulnerability detection in source code, designed to ensure there is efficiency in identifying potential weaknesses. The system employs a fine-tuned CodeBERT model, which leverages transformer-based natural language processing techniques to learn semantic patterns in source code. By deploying the model locally within a Dockerized environment, the solution ensures that source code remains entirely under organizational control, eliminating reliance on external servers. The methodology integrates real time scanning for multiple programming languages (Python, Java, and JavaScript), a web-based upload interface, and a core detection algorithm capable of classifying vulnerabilities based on learned patterns. Evaluation results demonstrate strong detection performance, with an average score of 92% for accuracy, precision, recall and F1-score respectively. The findings highlight the potential of on-premise AI-based systems to revolutionize vulnerability detection by combining accuracy and scalability.