FVD-DPM: Fine-grained Vulnerability Detection via Conditional Diffusion Probabilistic Models

Software vulnerabilities pose a significant threat to software security. Nevertheless, existing vulnerability detection methods still struggle to effectively identify vulnerabilities and pinpoint vulnerable statements. In this paper, we introduce FVD-DPM: a novel Fine-grained Vulnerability Detection approach via a conditional Diffusion Probabilistic Model. FVD-DPM formalizes vulnerability detection as a diffusion-based graph-structured prediction problem. Firstly, it generates a new fine-grained code representation by extracting graph-level program slices from the Code Joint Graph. Then, a conditional diffusion probabilistic model is employed to model the node label distribution in the program slices, predicting which nodes are vulnerable. FVD-DPM achieves both precise vulnerability identification (slice-level detection) and vulnerability localization (statement-level detection). We evaluate FVD-DPM on five collected datasets and compare it against nine state-of-the-art vulnerability detection approaches. Experimental results demonstrate that FVD-DPM significantly outperforms the baseline approaches across various evaluation settings.