ScholarWorks@중앙대학교

초록

Industry 4.0 has transformed industries by accelerating the adoption of artificial intelligence of things (AIoT); however, it has also led to security risks, such as data leakage. Existing data protection research focuses on network layers, whereas application-layer rating systems rely on subjective evaluations, limiting consistency and applicability. This study proposes a scalable and objective AIoT security rating framework that clarifies ambiguities in the five-question rating system of the Korean Intellectual Property Office and unifies fragmented managerial and technical rating systems. By leveraging large language models (LLMs), the framework integrates a security rating model based on 14 impact factors with automated questionnaire generation. A novel percentage-based measurable constraint ensures objectivity and consistency. The fine-tuned Llama 3.1 8B Instruct model, optimized via direct preference optimization, can enhance customization and question relevance. Results across 13 metrics, including G-Eval and Security G-Eval, highlighted its superiority in aligning questions with prompts, thereby improving specificity and clarity over existing LLMs. A user survey validated its effectiveness with a score of 4.1 out of 5 for correlation and answerability, supported by a Cronbach's alpha of 0.878. This study thus introduces a robust and practical AIoT security rating framework, particularly for the manufacturing, healthcare, and transportation domains, reducing subjective biases while improving applicability.

키워드