A proactive approach to software security using DCodeBERT for vulnerability management

Abstract

The complexity of modern software has increased security risks, emphasizing the need for automated detection and correction. DCodeBERT, a CodeBERT-based vulnerability detection and remediation framework, is introduced in this study. DCodeBERT uses a multi-task learning framework with shared-private layers, gradient normalization, and uncertainty weighting to stand out. This architecture lets the model capture general representations while preserving task-specific details. From open-source repositories and vetted vulnerability databases, 85,000 code snippets—vulnerable, clean, and repaired—were collected. C, C++, Java, and Python programming languages (PLs) make this dataset highly usable. DCodeBERT surpasses CodeGPT, VulDeePecker, CodeT5 Small, GraphCodeBERT, and Devign in accuracy, precision, recall, and F1-score. Statistics show that the improvements are significant, and qualitative inspection shows that the resulting patches fix buffer overflows and injection problems within semantic validity. This novel approach combines multi-task optimization with natural and PL semantic integration for high cross-language performance. The findings show that DCodeBERT improves vulnerability management in software development settings.