MIL 기반 생성형 AI 탐지 모델의 성능 및 해석 가능성 군사 보고서를 중심으로

Abstract

The widespread adoption of Generative Artificial Intelligence (Gen AI) has made it increasingly difficult to distinguish between human-written and Al-generated text, particularly in high-trust domains such as defense. This study proposes a detection framework based on Multiple Instance Learning (MIL) to enhance both performance and interpretability. Unlike traditional document-level approaches, the proposed model learns semantic relationships among sentences through sentence-level embeddings and an attention mechanism, enabling structural detection beyond surface linguistic cues. Experiments were conducted using 230 student reports from the Korea National Defense University, with self-reported AI usage scores (0-10) as primary labels. An additional 21 ground-truth (GT) documents were injected to evaluate the effect of label reliability. The MIL models outperformed the XGBoost baseline, showing an improvement of 9.24 percentage points in Macro FI (p < 0.001) and achieving AUC scores between 0.71 and 0.73, indicating statistically significant discriminative capability. GT injection stabilized decision boundaries and clarified semantic clustering in the embedding space. Qualitative analysis of attention weights provided interpretable evidence of model decisions, demonstrating explainable detection behavior. These findings highlight that limited but high-quality data can substantially improve both accuracy and transparency, offering a reliable and interpretable AI detection framework applicable to defense and other mission-critical domains.