Neural network structure and training method for detecting manipulated digital images

Abstract

We live in a digital image era. With digital cameras and mobile cameras, we can shoot digital images anytime, anywhere, and share images with social network services or mobile messenger services to anyone. When we look at digital images, we tend to believe that they are real scenes. In many cases, however, digital images contain fake information that has not happened. Today, there are more and more the manipulated images which contain fake information because anyone can make manipulated images easily by using image-editing programs such as Photoshop or by using deep learning method which called Deepfake. These manipulated images can be exploited in a variety of places, such as fake news, document forgery, and this leads to various social problems. To prevent these social problems researchers have proposed image manipulations detection methods in various aspects since about twenty years ago, however, most of the methods can be applied to only specific environments, and it is difficult to use them for real-world images. This paper proposes an image manipulation detection method that operates in an environment similar to real image distribution environment by using a neural network. In the second chapter, we first introduce the paired mini-batch training method, which allows neural networks to learn subtle differences of images faster and more accurately, and prove the efficiency of the training method through various experiments. Then we introduce the actual manipulated images obtained through image forensic service for about two years and describe the making process of a new double JPEG image dataset with using 1120 quantization tables extracted from the collected JPEG images in the third chapter. After that, we propose the new network for detecting double JPEG and show detecting various manipulations in the JPEG images using the trained network.