کنفرانس بین المللی فناوری اطلاعات و دانش

English

صفحه اصلی / شانزدهمین کنفرانس بین المللی فناوری اطلاعات و دانش

Robustness Gap in NLP Models for Vulnerability Descriptions: Benchmarking and Data Augmentation

نویسندگان :

AmirHossein Majd¹ Mahdi Yousefikia² Saghar Ghasemzadeh³ Amirreza Asari⁴ Arya Khoshnavataher⁵ Seyedeh Leili Mirtaheri⁶

1- University of Calabria 2- دانشگاه خوارزمی 3- دانشگاه خوارزمی 4- دانشگاه خوارزمی 5- دانشگاه خوارزمی 6- University of Calabria

کلمات کلیدی :

Software Vulnerabilities،Natural Language Processing،Robustness Benchmark،Noise Injection،Exploitability Prediction،Data Augmentation،Cybersecurity

چکیده :

Software vulnerability descriptions from CVE/NVD are the primary corpus for analysis, prioritization, and risk management in cybersecurity. Yet natural noise (typos, synonym substitutions, lexical variety) and adversarial perturbations undermine the accuracy and trustworthiness of NLP models. This paper presents, to our knowledge, the first systematic benchmark of NLP robustness on vulnerability descriptions. We train nine diverse architectures—lightweight transformers (MiniLM, MPNet, SBERT), hybrid models (BERT-LSTM, TextRCNN), and classical recurrent networks (BiLSTM, LSTM)—on a balanced dataset of over 56,000 real-world records from NVD and Exploit-DB, and fine-tune them for exploitability prediction. For comprehensive evaluation, we inject three noise families into test sets at levels from 10% to 80%: character-level edits (substitutions/swaps), synonym replacements using WordNet, and composite adversarial attacks generated with TextAttack. Performance declines across all models as noise rises, but vulnerability profiles differ: MiniLM attains the strongest clean-data score (F1 ≈ 0.933) yet is most brittle under character noise, whereas TextRCNN, despite a lower baseline, preserves comparatively higher stability in heavily perturbed conditions. Finally, we test a pragmatic hardening strategy—data augmentation with noisy variants followed by retraining—which consistently narrows robustness gaps across architectures without materially sacrificing clean-data accuracy. The benchmark and code enable reproducible evaluation and future robust modeling in cybersecurity.