On deep learning approach in remote sensing data forgery detection
Andrey V. Kuznetsov · 2020
Digital image forgery is a known issue due to the increasing availability of technologies and software that make it easy to create distorted images. In order to counter such attacks, several approaches have been developed to detect fakes. Of particular importance are the methods for detecting distortions of a particular type of digital image - Earth remote sensing data, which can be used to ensure the safety of protected areas, monitor the state of the environment, etc. This article proposes a new scheme based on neural networks and deep learning, which is based on the use of the new convolutional neural network (CNN) architecture to improve the quality of detection of the most common type of attack on digital images - embedding duplicates. As part of the proposed architecture, additional preprocessing layers are used to improve detection quality. This approach also demonstrates invariance to distortions introduced into the duplicated region. Experiments show that the proposed solution exceeds the known copy-move detection algorithms - the metric value F1 reaches 0.77. At the same time the proposed deep learning approach shows high quality for the splicing detection task.