Multi-level Data Hiding for Digital Image and Video Min Wu a b* Hong Heather Yu b Alex Gelman b ** a Electrical Engineering Dept., Princeton Univ., Princeton, NJ 08544 b Panasonic Information & Networking Technologies Laboratory, Princeton, NJ 08540 ABSTRACT Previous works on data hiding generally targeted on a specific tradeoff between capacity and robustness. This results in overestimation of the processing noise under some situations and/or underestimation under some other situations, hence limits the overall performance. In this paper, we propose a multi-level data hiding scheme which is able to convey secondary data in high rate when noise is not severe and can also convey some data reliably under heavy distortion. The proposed scheme is motivated by a two-category classification of embedding schemes and by a study on detection performance of spread spectrum watermarking. The multi-level data hiding has been successfully applied to both digital image and video, and can be used for applications such as copy control. Keywords: data hiding, digital watermarking, copy control. 1. INTRODUCTION In the digital world, we often deal with two types of data: the traditional ones and the perceptual ones. For traditional data, such as the ASCII files and executable codes, losing or changing a single bit may make the whole file useless. On the other hand, perceptual data such as digital audio, image, and video, can experience minor changes without causing noticeable difference. Such an interesting property makes it possible to hide secondary data in host media for the purpose of identification, annotation, copy control, and copyright protection. This paper emphasizes on the issues of embedding robustly and imperceptibly a non-trivial number of bits in high-quality digital image and video. As has been pointed out in literature [1], data hiding can be thought as a special communication problem, where secondary data are the information to be sent from sender to receiver through a special channel. The channel is compromised of the host signal (image or video) and the noise introduced by signal processing and/or attacks. Imperceptibility, robustness against moderate compression and processing, and the ability to hide many bits are the basic but rather contradictory requirements for many data hiding applications. The traditional way to handle this is to target at a specific capacity-robustness pair. Some approaches choose to robustly embed just one or a few bits [2][3], while others choose to embed a lot of bits but to tolerate little or no distortion. However, a single robustness threshold generally overestimates the noise condition in some situations and/or underestimates in some other situations. It is desirable to design a data hiding system that is able to convey secondary data in high rate when noise is not severe and can also convey some data reliably under severe processing. This paper proposes multi-level data hiding, which can achieve the above goal. The proposed scheme is motivated by a two-category classification of embedding schemes and by a study on the performance of non-coherent detection of the popular spread spectrum watermarking [4][5][6]. We shall discuss the proposed scheme and provide the reasoning and analysis in Section 2. Experimental results of hiding data in digital images and videos are shown in Section 3. We summarize and present the direction of future works in the final section. * This work was performed while the author was a summer intern at Panasonic Information & Networking Technologies Laboratories (PINTL). ** The authors can be contacted at minwu@ee.princeton.edu, {heathery, alex}@research.panasonic.com.