Revisiting the Performance of Deep Learning-Based Vulnerability Detection on Realistic Datasets
Bibliographic record
Abstract
The impact of software vulnerabilities on everyday software systems is concerning. Although deep learning-based models have been proposed for vulnerability detection, their reliability remains a significant concern. While prior evaluation of such models reports impressive recall/F1 scores of up to 99%, we find that these models underperform in practical scenarios, particularly when evaluated on the entire codebases rather than only the fixing commit. In this paper, we introduce a comprehensive dataset (Real-Vul) designed to accurately represent real-world scenarios for evaluating vulnerability detection models. We evaluate DeepWukong, LineVul, ReVeal, and IVDetect vulnerability detection approaches and observe a surprisingly significant drop in performance, with precision declining by up to 95 percentage points and F1 scores dropping by up to 91 percentage points. A closer inspection reveals a substantial overlap in the embeddings generated by the models for vulnerable and uncertain samples (non-vulnerable or vulnerability not reported yet), which likely explains why we observe such a large increase in the quantity and rate of false positives. Additionally, we observe fluctuations in model performance based on vulnerability characteristics (e.g., vulnerability types and severity). For example, the studied models achieve 26 percentage points better F1 scores when vulnerabilities are related to information leaks or code injection rather than when vulnerabilities are related to path resolution or predictable return values. Our results highlight the substantial performance gap that still needs to be bridged before deep learning-based vulnerability detection is ready for deployment in practical settings. We dive deeper into why models underperform in realistic settings and our investigation revealed overfitting as a key issue. We address this by introducing an augmentation technique, potentially improving performance by up to 30%. We contribute (a) an approach to creating a dataset that future research can use to improve the practicality of model evaluation; (b)Real-Vul– a comprehensive dataset that adheres to this approach; and (c) empirical evidence that the deep learning-based models struggle to perform in a real-world setting.
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 | 0.007 | 0.024 |
| Meta-epidemiology (narrow) | 0.003 | 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 | 0.004 |
| Open science | 0.003 | 0.002 |
| Research integrity | 0.003 | 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".