基於多種深度學習語言模型之CVSS危險等級預測準確度提升之研究

Accuracy Comparison between LSTM and BERT Models Based on Multi-Item Classification–Take Movie Reviews as an Example

在當前網路環境中，駭客攻擊的頻率與複雜性不斷提升，對全球企業的資訊安全構成嚴峻挑戰。為了有效因應這些日益嚴重的威脅，企業與組織紛紛投入資源以強化其資安防禦能力。由於攻擊手法日新月異且技術持續演進，企業急需一套標準化的系統，來準確評估與管理各類資安風險。
本研究旨在探討如何基於CVSS評分規則，預測資安事件的CVSS危險等級評分（即CVSS基礎分數），以協助企業更精確地判斷事件的嚴重性。CVSS透過評估漏洞的攻擊向量、攻擊複雜度、權限要求、使用者互動、機密性影響、完整性影響及可用性影響等多項指標，提供企業一套詳細且具一致性的評分框架，有助於資源優先排序與風險控制。
研究蒐集GitHub上CVE Program所公布之2020至2024年間之JSON格式資料，進行欄位解析與資料清洗後，建立一套以深度學習為核心的CVSS危險等級評分預測系統。模型部分分別採用RoBERTa與DeBERTa自然語言模型進行回歸任務訓練，並透過不同超參數組合進行效能比較。最後結合資料視覺化模組，將模型預測結果與實際標註值進行圖像比對，使模型效能差異一目了然，有助於企業在事件初期快速掌握風險等級，進一步完善資安決策流程。

In today’s increasingly complex and interconnected network environment, the frequency and sophistication of cyberattacks are steadily rising, posing significant challenges to global information security. To effectively address these growing threats, enterprises and organizations are investing considerable resources to enhance their cybersecurity defenses. Given the rapidly evolving nature of attack techniques and technologies, there is an urgent need for a standardized system that can accurately assess and manage various types of cybersecurity risks.
This study aims to explore how to predict the CVSS (Common Vulnerability Scoring System) base scores of cybersecurity incidents based on CVSS scoring criteria, thereby assisting organizations in more accurately evaluating the severity of such incidents. CVSS provides a detailed and consistent scoring framework by assessing multiple indicators, including attack vector, attack complexity, privileges required, user interaction, and the impacts on confidentiality, integrity, and availability. This enables organizations to better prioritize resources and implement effective risk control strategies.
The research collects JSON-formatted data published by the CVE Program on GitHub from 2020 to 2024. After parsing and cleaning the dataset, a CVSS risk score prediction system based on deep learning is constructed. The study employs RoBERTa and DeBERTa—two natural language processing models—for regression tasks, and compares their performance using various hyperparameter configurations. Finally, a data visualization module is integrated to visually compare the model predictions with actual CVSS base scores, clearly highlighting performance differences. This approach helps enterprises quickly assess risk levels at the early stages of incidents and improve their cybersecurity decision-making processes.