AMBER: A Semantic Aware Detection of Code Vulnerabilities and Their Clones
TL;DR
A semantic-aware framework that utilizes semantic embeddings, structural program features, and similarity metrics to construct a multi-channel representation, which is fed to a convolutional neural network for vulnerability detection, and then employed in a FAISS-based semantic similarity search for finding vulnerable code segments that can be reused in other large codebases.
Abstract
The identification of vulnerabilities in software and their clones is crucial for secure software development, but achieving high accuracy whilst maintaining a scalable detection system is difficult. We introduce Amber, a semantic-aware framework that utilizes semantic embeddings, structural program features, and similarity metrics to construct a multi-channel representation, which is fed to a convolutional neural network (CNN) for vulnerability detection, and then further employed in a FAISS-based semantic similarity search for finding vulnerable code segments that can be reused in other large codebases. In addition to real-world data from open-source repository data sets, Amber is assessed on 12,303 vulnerable functions and 21,057 non-vulnerable functions from the SARD vulnerability database. Regarding accuracy, recall, and F1-score, Amber demonstrated superior performance over several leading baselines, including VulCNN, VulDeePecker, SySeVR, and Devign. In addition, Amber enables post-classification vulnerable clone detection, providing a deep understanding of vulnerabilities through hidden and subtle vulnerable clones. This work provides a scalable and explainable framework for development and security practitioners, helping to advance the automation of vulnerability and their clone detection and improve software security practices overall.