Abstract
When deepfakes are widespread on chatting platforms, they are expected to be subject to heavy resizing and compressing steps. In this paper, we present a tensor-based representation of compressed and resized images. Tensor embeds DCT features computed on multi-scaled and multi-compressed versions of the input facial image. Moreover, a custom deep-architecture is designed and trained on the proposed representation. Experimental results show its pros and cons with respect to state-of-the-art methods.
AUTHORS Sara Concas; Gianpaolo Perelli; Gian Luca Marcialis; Giovanni Puglisi
DOI 10.1109/ICIP46576.2022.9897606
Published in: 2022 IEEE International Conference on Image Processing (ICIP)