설명가능 인공지능 기반 중요 얼굴 영역 탐색을 통한 효율적인 FAS(Face Anti-Spoofing) 모델 구축

Abstract

Compared to other biometric methods, facial recognition is relatively slow and vulnerable. To address this issue, the development of fast and accurate Face Anti-Spoofing (FAS) models is essential. In this study, we propose an explainable neural network-based approach that leverages important facial areas to construct an efficient FAS model. These important areas are quantitatively identified by analyzing the prediction mechanism of the FAS model. To validate the proposed approach, we train a new model using images that include only the identified important areas and compare its performance to that of the traditional model. The results demonstrate that the performance of the two models is comparable, indicating the feasibility of replacing existing models. Additionally, the proposed method reduces computational overhead by pre-removing irrelevant areas, enabling the construction of an efficient FAS model that focuses on learning from the key facial areas.