생성형 AI의 데이터 학습 여부 판별을 위한 다층적 보안 검증 프레임워크에 관한 탐색적 연구

An Exploratory Study on a Multi-layered Security Verification Framework for Identifying Data Training in Generative AI

초록

The recent proliferation of generative AI, particularly Large Language Models (LLMs), has significantly enhanced business efficiency across various sectors. However, it also introduces a novel security risk: irreversible information leakage when input data is incorporated into model training. While conventional information security frameworks focus on data protection during transmission and storage, they possess fundamental limitations in generative AI environments where data becomes internalized within model weights and may be reproduced even after its deletion. To address these challenges, this study categorizes generative AI security threats into 'Log-based' and 'Learning-based' leakage, conducting an exploratory research focused on proposing a multi-layered security verification framework for the latter. We systematically examine key verification methodologies—including Perplexity-based analysis, canary tokens, Membership Inference Attack (MIA), and Gradient Analysis—across both black-box and white-box environments, and design step-by-step verification protocols and decision logics for practical implementation. The exploratory analysis reveals that a single technique is insufficient for completely verifying data training in generative AI, necessitating a hybrid verification strategy tailored to the verifier's access privileges. Specifically, this research proposes a dual-track verification system where statistical indirect methods are effective for general users in API-based environments, while direct verification techniques like Gradient Analysis provide deterministic evidence for auditors in white-box environments. The significance of this study lies in transitioning the generative AI security paradigm from simple data protection to a verifiable problem and establishing a foundation for future empirical research through exploratory analysis. This provides a technical basis for securing data sovereignty during the adoption of generative AI and serves as a fundamental resource for designing verification-centric AI governance and Machine Unlearning policies.

키워드

생성형 인공지능; 대규모 언어모델; LLM; 데이터 학습 여부 검증; 멤버십 추론 공격; 퍼플렉시티 분석; Generative AI; Large Language Models; LLMs; Data Learning Verification; Membership Inference Attack; Perplexity Analysis
제목
생성형 AI의 데이터 학습 여부 판별을 위한 다층적 보안 검증 프레임워크에 관한 탐색적 연구
제목 (타언어)
An Exploratory Study on a Multi-layered Security Verification Framework for Identifying Data Training in Generative AI
저자
박설아; 장항배
DOI
10.7838/kjdi.2026.31.3.151
발행일
2026-08
유형
Y
저널명
한국디지털산업학회지
권
31
호
3
페이지
151 ~ 172