Interpretable Artificial Intelligence Models Based on Large Language Models for Web Application Vulnerability Detection and Security Analysis
Year: 2026
Type: conference-paper
Abstract: Web application security remains difficult because code complexity and attack techniques evolve faster than traditional rule-based analysis. Classical static analysis is limited by false positives, rule dependency, and weak semantic understanding, while many Machine Learning and Deep Learning detectors still provide limited interpretability. This paper presents a survey and experimental evaluation of interpretable Large Language Model-based vulnerability detection for web applications. The survey analyzes peer-reviewed publications on transformer architectures, hybrid models, and explainability techniques. The experiment evaluates three recent generative models on curated web vulnerability snippets across five vulnerability categories. It measures detection accuracy and explanation quality through vulnerability type identification, vulnerable line localization, and fix suggestion generation. The results show that the strongest model achieves the best detection performance, while structured reasoning produces the strongest explanation quality. The study provides empirical evidence toward accurate and transparent Artificial Intelligence-driven security analysis for web applications.
Institution: Astana IT University
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7203606476