An Improved Template Inversion Attack Against Korean Face Images

Recent template inversion methods achieve strong performance on Western datasets, but poor results on Korean facial images. This gap poses security risks for the authentication system using facial recognition in Korea, as the stakeholders may misunderstand that protecting facial templates on their databases is not necessary. We propose an enhanced template inversion method with three improvements: refined preprocessing using per-eye alignment and GFPGAN upscaling, MSE-enhanced loss function, and dynamic weight adjustment. Our method outperforms existing approaches, achieving 0.882 cosine similarity, 9.7 L₂ norm, and 0.312 LPIPS. Beyond reconstruction quality, we evaluate the security implications of template inversion attacks by analyzing verification robustness using Successful Attack Rate (SAR) at fixed False Match Rate (FMR) operating points, Receiver Operating Characteristic (ROC) curves, and Equal Error Rate (EER), providing a comprehensive assessment under realistic authentication thresholds. These results emphasize the critical need for robust facial template protection in facial recognition authentication systems.