오토인코더 기반 재현 이미지 안전성 평가

다국어 초록 (Multilingual Abstract)

This study aims to establish evaluation criteria for the safety of synthetic image data by proposing a method to quantitatively assess the risk of privacy infringement. Synthetic images that are nearly identical to the originals due to overfitting or one-to-one transformations cannot be considered safe. However, detecting such unsafe images is not trivial. To address this, we propose a feature-based similarity detection method that leverages an autoencoder along with SSIM, LPIPS, and cosine similarity. Furthermore, we introduce a distribution-based approach that uses the original data’s characteristics to determine the safety of synthetic images.