Study on detecting subtle signals in images using deep neural networks

Abstract

With the development of deep neural networks (DNN), the performances of image classification, object detection, and image segmentation have been improved unprecedentedly. However, the improvements are usually achieved in the field of high-level computer vision tasks, not in the case of understandings of low-level signals in images. In this dissertation, we propose DNN-based methods for identifying subtle signals in images, specifically, for steganalysis and double JPEG compression detection. Image steganalysis refers to the technique classifying the normal cover image and stego image where messages are embedded by steganography. As steganography aims to hide the existence of messages in images, the traces induced by message embedding are very subtle. Therefore, well-defined convolutional neural networks (CNN) for tackling high-level vision tasks cannot train the differences between cover and stego. CNN-based steganalysis methods usually include constraints, such as a high-pass filter to suppress the image contents, to make CNN trainable. We present an end-to-end trainable CNN with normalization-free residual modules. It shows state-of-the-art performance in spatial and JPEG steganalysis. Image steganalysis ideally aims to identify stego images embedded by unknown steganography. Although CNN-based steganalysis improved their performance against single known steganography, it is far from practical scenarios. We present two methods for improving practicality for JPEG steganalysis. Firstly, we present a multi-class steganalysis, which classifies three steganography of three different JPEG quality factors. Next, we present resource-efficient CNN-based steganalysis taking RAW DCT coefficients as inputs for JPEG images of various resolutions. Double JPEG compression detection is essential in that it can be used to identify image manipulations regardless of manipulation types. Although existing double JPEG networks show good performances, they require preprocessing such as constructing DCT histograms. We present an end-to-end trainable CNN for identifying double JPEG compression by taking RAW DCT coefficients and achieved state-of-the-art performances.