Anti-forensics of Image Sharpening Using Generative Adversarial Network

Abstract

Image sharpening is an image enhancement method which has been widely used to improve the quality of images. Therefore, in image forensics, it is required to be identified as all possible manipulations applied in images need to be detected. In recent years, sharpening detection get evolved with new detectors proposed every year to gradually boost the detection performance. This situation continues for several years till the introduction of convolutional neural networks (CNNs). With the assistance of CNNs, the detection of sharpening seems to be completely solved that the detection performance for sharpening achieves perfect, even when the images are weakly sharpened. Is it true that we should no longer pay attention to sharpening forensics any more? To answer this question, in this paper, an anti-forensics method based on generative adversarial network (GAN) is proposed to investigate the philosophy. The images generated via our method possess the feature of sharpening, however, they cannot be simply considered as sharpened images because no traditional sharpening manipulation is applied during the procedure. Observed from the experimental results, even the state-of-art sharpening detector based on CNN can be deceived with the GAN generated images.