Deep Learning-Based Detection of Code Execution Vulnerabilities in Binary Programs

Deep Learning-Based Detection of Code Execution Vulnerabilities in Binary Programs

연구 분야: Strategies

학회: International Conference on Neural Information Processing

Numerous incidents have shown that code execution vulnerabilities in binary files pose significant security risks, as attackers exploit them to compromise system integrity. This issue has garnered extensive attention from academia and industry. Binary code vulnerabilities impact both traditional and emerging software domains. Smart contracts, deployed in blockchain environments, face similar challenges. They must be compiled into machine-readable binary code before execution and are exposed to open networks, making them as frequent targets of attacks. Traditional vulnerability detection techniques rely on manual analysis or rule-based heuristics, which are time-consuming and inapplicable to detecting complex vulnerabilities. This study proposes a new approach using deep learning models to automatically detect code execution vulnerability paths in binary programs. Experiments demonstrate the method’s feasibility, achieving an accuracy rate of 97%–98% on the BERT+RNN model.