Graph dataset - LibTiff

A comparative, empirical study of state-of-the-art contrastive and generative graph learning models applied to source and binary software fragments drawn from the National Vulnerability Database (NVD) reveals that Graph Masked Auto-Encoders show exceptional promise for detecting security vulnerabilities, outperforming all other baseline models in the study. This fills a key gap in the literature on automated and machine-assisted discovery and patching of software security vulnerabilities, which has become increasingly critical with the dramatic increase in modern software complexity, but for which Graph Neural Network (GNN) approaches are understudied relative to traditional processes, such as manual source code auditing and fuzzing.To conduct the study, a novel dataset is first collected by extracting vulnerable code fragments from six applications with NVD-documented security flaws and converting these codes to five different graph types using specialized tools based on code property graphs and binary semantics lifting. The resulting dataset is applied to GNN-based analyses to determine which algorithm and graph type performs best, followed by an ablation study to determine which combination of parameters maximizes effectiveness of the top-performing detector. The study is the first to train GNN models on a combination of source- and binary-level code features, which is important for helping cyber defenders craft source-level patches that defend against binary-level attacks.