(A) study on multimedia security for various types of media

Abstract

With the development in media sharing platform and mobile device equipped with a camera, the era has come in which media such as image and video can be easily generated via smartphone and shared through social network services and mobile messengers. However, these media-related advances have benefited the lives of modern people and, at the same time, caused various security problems for media contents. In particular, copyright problem about media ownership, which is caused by illegal redistribution, and social problem due to content modified through editing tools such as Photoshop are frequently issued. To deal with these issues, research on multimedia security has been conducted, and multimedia security is generally categorized into the following two items: digital watermarking and multimedia forensics. Digital watermarking is an approach to protect copyrighted works, which covertly inserts noise-like watermark for copyright ownership into a media, and multimedia forensics aims at authenticating integrity of various types of media including images and videos by detecting artifacts caused by manipulations. In this dissertation, four novel approaches to multimedia security for various types of media are proposed: two for digital watermarking and two for multimedia forensics. With regard to digital watermarking, we first propose non-subsampled contourlet transform (NSCT)-based robust and perceptual watermarking for depth-image-based rendering (DIBR) 3D Images. Through extensive experiments, our watermarking framework showed stable and high watermark extraction performance against DIBR attacks, signal processing operations, and geometric distortions, while maintaining high imperceptibility. In addition, to evaluate the robustness of existing conventional multi-bit watermarking methods for 2D images and DIBR 3D images, we introduce convolutional neural network (CNN)-based watermarking attack network (WAN). The proposed WAN can learns the weak points of the targeted watermarking approach and attacks watermarked images to mislead the watermarking extractor with minimal visual degradation. With regard to multimedia forensics, we first propose a CNN-based approach to classify artifacts caused by seam carving-based image retargeting for both reduction and expansion. After that, two-stream network for capturing low-level forensic artifacts of I-frame and P-frame in double-compressed H.264 videos is proposed. Compared to comparative methods, our forensics approaches shows the state-of-the-art performance in terms of manipulation classification.