VulDeNoise: Outlier Detection to Reduce Label Noises for Effective Vulnerability Detection

While automated vulnerability detection approaches, especially those using Graph Neural Networks (GNNs), have shown remarkable promise, their effectiveness is often constrained by significant label noise in prevalent vulnerability datasets. To address this, we propose VulDeNoise , an innovative dataset denoising framework. Our approach is grounded in multi-view learning theory, postulating that correctly labeled samples exhibit consistent training dynamics across different code graph representations, whereas mislabeled samples manifest significant discrepancies due to inherent semantic conflicts. VulDenoise operationalizes this by training a target detector on each of the three code representation graphs, constructing a loss vector for each sample from its cross-view training loss sequences, and then employing an ensemble of outlier detection algorithms to find noisy instances. We conducted extensive experiments on four prominent GNN-based detectors ( Devign , DeepWukong , ReVEAL , and IVDetect ) using the Big-Vul and FFmpeg+QEMU datasets. The results demonstrate that training on Big-Vul dataset denoised by VulDeNoise consistently enhances the F1-score of these four detectors by 5-10%. In controlled experiments on the FFmpeg+QEMU dataset, where label noise was synthetically added at varying ratios, VulDeNoise achieved a denoising F1-score of up to 70%, demonstrating its high effectiveness in identifying noisy labels. Furthermore, VulDeNoise substantially outperforms state-of-the-art denoising methods like Confident Learning and Differential Training , and even surpasses a Large Language Model (LLM)-based auditing approach. Ablation studies confirm the robustness of our design, revealing that the synergy of all three code representations and a carefully selected training duration are essential for optimal performance. VulDeNoise offers an effective, automated solution for improving the quality of vulnerability datasets, paving the way for more reliable deep learning-based vulnerability detection models.