Stream/Area:
-
Subject:
-
Paper ID:
IJSREM69139
Volume:
10 |
11 |
Nov 2026
Pages:
1-21
Publication Frequency:
Monthly (12 issues per year)
License:International Journal of Scientific Research in Engineering and Management follows the Creative Commons Attribution 4.0 (CC BY 4.0) License, allowing free use and distribution with proper attribution.

Software Engineering
Gandhinagar University

Software Engineering
Gandhinagar University
Source code vulnerabilities remain a leading cause of security incidents, yet automated detection is still unreliable. Static analyzers such as CodeQL and Semgrep are precise on known patterns but miss logic-dependent flaws and generate many alerts that developers must triage. Fine-tuned transformer models such as CodeBERT and GraphCodeBERT report high scores on public benchmarks, but recent work shows that duplicated samples, noisy labels and train–test leakage inflate these scores and that performance collapses on cleaner data. Prompted large language models (LLMs) can explain their reasoning, but their answers are non-deterministic, they over-report vulnerabilities, and they sometimes justify correct verdicts with incorrect reasoning. This paper proposes a hybrid, verification-grounded framework. A static-analysis stage and program slicing supply context, a fine-tuned code encoder supplies a prior, and an open-source LLM produces a structured finding (CWE type, evidence lines and rationale). A verification layer then tests the LLM’s claim against the code through evidence-line grounding, static data-flow support and self-consistency across sampled runs, and a logistic fusion of these signals yields a calibrated confidence score. The contribution is the pipeline design and a leakage-aware evaluation protocol, not a new model architecture. We specify the datasets (Big-Vul, Devign, PrimeVul and Juliet), baselines, metrics, ablations and four testable hypotheses. Experimental results are not yet available and are reported as placeholders pending implementation.
KEYWORDS
This paper proposed a verification-grounded hybrid framework for source-code vulnerability detection. It combines static-analysis context, a fine-tuned GraphCodeBERT prior and a structured LLM finding, and it checks the LLM’s claims through evidence grounding, static corroboration and sampled-run stability before producing a calibrated confidence. The paper also specified a leakage-aware evaluation protocol covering Big-Vul, Devign, PrimeVul and Juliet, a set of baselines and ablations, and four hypotheses. The contribution is the design and the evaluation plan, and no performance claims are made until the experiments are completed. Future work includes extending the context builder to repository-level analysis, learning the fusion function from richer features, testing larger and more recent open models, adding automated patch suggestions that are validated by re-running the detectors, and evaluating robustness to semantics-preserving code changes of the kind that disrupted LLMs in earlier work [12]. A user study on whether confidence scores and evidence presentation reduce developer triage time would also test the practical value of the approach.
[1] MITRE, “Common Weakness Enumeration (CWE).” [Online]. Available: https://cwe.mitre.org/ (accessed: [date to be added]). [2] B. Chess and G. McGraw, “Static analysis for security,” IEEE Security & Privacy, vol. 2, no. 6, pp. 76–79, 2004, doi: 10.1109/MSP.2004.111. [3] B. Johnson, Y. Song, E. Murphy-Hill, and R. Bowdidge, “Why don’t software developers use static analysis tools to find bugs?,” in Proc. 35th Int. Conf. Software Engineering (ICSE), 2013, pp. 672–681, doi: 10.1109/ICSE.2013.6606613. [4] B. P. Miller, L. Fredriksen, and B. So, “An empirical study of the reliability of UNIX utilities,” Commun. ACM, vol. 33, no. 12, pp. 32–44, 1990, doi: 10.1145/96267.96279. [5] R. Russell et al., “Automated vulnerability detection in source code using deep representation learning,” in Proc. 17th IEEE Int. Conf. Machine Learning and Applications (ICMLA), 2018, arXiv:1807.04320. [6] Z. Li et al., “VulDeePecker: A deep learning-based system for vulnerability detection,” in Proc. Network and Distributed System Security Symp. (NDSS), 2018, doi: 10.14722/ndss.2018.23158. [7] Y. Zhou, S. Liu, J. Siow, X. Du, and Y. Liu, “Devign: Effective vulnerability identification by learning comprehensive program semantics via graph neural networks,” in Advances in Neural Information Processing Systems (NeurIPS), vol. 32, 2019, arXiv:1909.03496. [8] Z. Feng et al., “CodeBERT: A pre-trained model for programming and natural languages,” in Findings of the Association for Computational Linguistics: EMNLP 2020, 2020, doi: 10.18653/v1/2020.findings-emnlp.139. [9] M. Chen et al., “Evaluating large language models trained on code,” arXiv:2107.03374, 2021. [10] B. Rozière et al., “Code Llama: Open foundation models for code,” arXiv:2308.12950, 2023.
Publication Date
11/18/2026
Peer Review Type
Double-Blind Peer Review
Review Rounds: 2
Documents
Article Views
0
Total views since publication
PDF Downloads
0
Total downloads since publication
Geographic Reach
5
Digital Preservation: This article is permanently archived on Zenodo and complies with all data preservation standards across major indexing platforms.
Plagiarism Screening
This manuscript has been checked for plagiarism using TurnitinPlagiarismChecker software.
Conflict of Interest
The authors declare no conflict of interest related to this research work.
Funding Information
No funding information available for this research.
Related research in the same subject area
Loading similar articles...