HardVD: High-capacity cross-modal adversarial reprogramming for data-efficient vulnerability detection

Abstract

The substantial proliferation of software vulnerabilities poses a persistent threat to system security, driving increased interest in applying deep learning (DL) for vulnerability detection. However, DL-based detectors often operate with a fixed number of input tokens, leading to semantic loss over large code snippets. Additionally, developing these detectors demands substantial labeled data and training time. To address these limitations, this paper proposes HardVD, which explores High-capacity cross-modal adversarial reprogramming for data-efficient Vulnerability Detection. HardVD devises a high-capacity semantic extractor to capture salient features per line of code, which are then arranged as patches to form an image representing the target function. These images are processed using convolutional filters as universal perturbations and non-parametric label remapping to adapt a pretrained Vision Transformer (ViT) for vulnerability detection, updating only the limited parameters of the perturbation filters during training. Extensive experiments demonstrate that HardVD outperforms DL-based baselines in terms of detection effectiveness, data-limited performance, and computational overhead. The ablation study also confirms the essential role of our high-capacity semantic extractor, without which an averaged relative decrease of 5.87% and 7.98% in accuracy and F1 score is observed, respectively.