VGX: Large-Scale Sample Generation for Boosting Learning-Based Software Vulnerability Analyses

TL;DR

VGX tackles the scarcity and noise in vulnerability data for learning-based defenses by combining semantics-aware contextualization with human-knowledge-enhanced edit-pattern formation. It uses a value-flow-informed Transformer with specialized pre-training and a curated set of vulnerability-edit patterns to inject realistic vulnerabilities into normal code at scale. The approach yields a large, high-quality vulnerability dataset (around 150K samples) and demonstrably improves downstream tasks such as vulnerability detection, localization, and repair, while also enabling discovery of real-world CVEs. This work provides practical impact by offering open-source code and datasets to advance defensive software vulnerability analyses and related research.

Abstract

Accompanying the successes of learning-based defensive software vulnerability analyses is the lack of large and quality sets of labeled vulnerable program samples, which impedes further advancement of those defenses. Existing automated sample generation approaches have shown potentials yet still fall short of practical expectations due to the high noise in the generated samples. This paper proposes VGX, a new technique aimed for large-scale generation of high-quality vulnerability datasets. Given a normal program, VGX identifies the code contexts in which vulnerabilities can be injected, using a customized Transformer featured with a new value-flowbased position encoding and pre-trained against new objectives particularly for learning code structure and context. Then, VGX materializes vulnerability-injection code editing in the identified contexts using patterns of such edits obtained from both historical fixes and human knowledge about real-world vulnerabilities. Compared to four state-of-the-art (SOTA) baselines (pattern-, Transformer-, GNN-, and pattern+Transformer-based), VGX achieved 99.09-890.06% higher F1 and 22.45%-328.47% higher label accuracy. For in-the-wild sample production, VGX generated 150,392 vulnerable samples, from which we randomly chose 10% to assess how much these samples help vulnerability detection, localization, and repair. Our results show SOTA techniques for these three application tasks achieved 19.15-330.80% higher F1, 12.86-19.31% higher top-10 accuracy, and 85.02-99.30% higher top-50 accuracy, respectively, by adding those samples to their original training data. These samples also helped a SOTA vulnerability detector discover 13 more real-world vulnerabilities (CVEs) in critical systems (e.g., Linux kernel) that would be missed by the original model.

Figures (6)

• Figure 1: A motivating example on vulnerability injection.
• Figure 2: Design overview of the VGX approach, highlighting its two main phases: learning/training and production.
• Figure 3: The vulnerability type (CWE) distribution of generated vulnerable samples in the large-scale production.
• Figure 4: An example where VGX correctly predicts the statement at Line 8 (marked as cyan), but without value-flow-based position encoding it would incorrectly locates Line 2 (marked as yellow).
• Figure 5: An example where VGX successfully remove the statement at Line 6 (marked as cyan) to inject a memory leak vulnerability, but without manual pattern refinement via regex it would not be able to do so.