CODE-SEMANTIC BASED INTELLIGENT VULNERABILITIES DETECTOR

Abstract

The presence of vulnerabilities in a software system can significantly compromise its security, underscoring the importance of automatic software vulnerability detection. Current detection techniques, however, grapple with issues such as dependency problems, lexical bias, and low detection resolution. Though machine learning has been widely studied for predictive applications, its use in vulnerability detection has been hampered by lack of semantic information in deep learning techniques, which results in diminished accuracy. The study proposes an innovative automatic vulnerability detection system that synthesizes feature selection and deep neural networks (BiLSTM) to mitigate the challenges. This system selects features based on novel code metric that incorporates control flow and suspected functions, with feature selection implemented using a decision tree. The model is trained using standardized dataset (SARD) and is capable of detecting improper input validation, SQL injection attacks, cross-site scripting attacks, missing authorizations, and buffer overflow attacks. We compare our system's performance in terms of precision, recall, accuracy, and area under the receiver operating characteristics curve (AUC-ROC) with various techniques and commercial systems. The empirical evaluation demonstrates that our proposed system achieves a 4% improvement over commercial systems and a 6% improvement over other machine learning-based systems.