딥페이크 이미지 조작 기법과 학습 가능한 컨볼루션 커널을 통한 초고해상화 탐지

초록

In this paper, we propose a novel detection model to distinguish between original images and super-resolution (SR) images generated by both GAN-based algorithms (e.g., SRGAN, EDSR, RCAN) and a commercial tool (TOPAZ Labs). Previous deepfake detection studies have attempted to improve performance by leveraging frequency-domain characteristics or pre-trained backbone networks; however, they often suffered from limited adaptability, excelling only on specific algorithms or datasets. To address this, we introduce a framework that extracts fine-grained, local features of super-resolution images via learnable convolution kernels (LCKs) and leverages a ResNet-50 model pre-trained to robustly handle various image manipulations. Our approach demonstrates exceptional detection accuracy across diverse SR datasets, achieving a 100% success rate on all test sets. Furthermore, gradient map visualizations reveal that the proposed model effectively focuses on critical image details, such as edges and textures, to discriminate super-resolution images from their originals.

키워드