Evaluating Deep Learning-based Vulnerability Detection Models on Realistic Datasets
Bibliographic record
Abstract
The impact of software vulnerabilities on daily-used software systems is alarming. Despite numerous proposed deep learning-based models to automate vulnerability detection, the detection of software vulnerabilities remains a significant issue. While some techniques report high precision/recall scores of up to 99%, our experience leads us to believe that these models may underperform in realistic settings, specifically when evaluating vulnerability detection models on the entire source code repository of a project. Therefore, in this thesis, we create a more comprehensive vulnerability detection dataset (i.e., Comp-Vul), which aims to accurately represent the realistic settings where vulnerability detection models are deployed. Then, we evaluate the performance of two state-of-the-art deep learning-based models, LineVul and DeepWukong, on the Comp-Vul dataset. Our results show that the performance of both models drops drastically, with precision dropping by 86% - 95% and F1 score dropping by 88% - 91%. Our further investigation shows that the ratio of vulnerable to non-vulnerable samples in the evaluation dataset significantly impacts the performance metrics of these models. When we visualize the embeddings produced by the models, we find that there is a substantial overlap between vulnerable and non-vulnerable samples. This shows that these models have difficulty distinguishing between vulnerable and non-vulnerable samples in the Comp-Vul dataset, resulting in a high number of false positives. We introduce a new program slice-level vulnerability detection technique named SliceVul, which leverages the powerful capabilities of Transformers and incorporates the semantic properties of source code programs such as data and control flow information. Our approach outperforms the existing state-of-the-art program slice-level vulnerability detection model, DeepWukong when evaluated on the Comp-Vul dataset. Our study argues that accurately identifying vulnerabilities using deep learning remains a challenging task that requires improved approaches to model evaluation and design. Further research and development, complemented by realistic evaluation datasets, is required to enhance the performance of these methods.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.006 | 0.018 |
| Meta-epidemiology (narrow) | 0.004 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.003 | 0.002 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.002 | 0.003 |
| Open science | 0.003 | 0.002 |
| Research integrity | 0.002 | 0.003 |
| Insufficient payload (model declined to judge) | 0.001 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".