MULTI-LEVEL WATERMARKING WITH INDEPENDENT DECODING M. Butman Bar-Ilan University Department of Math and Computer Science Ramat-Gan 52905, Israel H.Z. Hel-Or Department of Computer Science Haifa University Haifa 31905, Israel ABSTRACT A drawback of most watermarking techniques is the need for some additional information in order to retrieve the wa- termark. Additionally, the robustness of the watermark de- creases as the number of information bits stored in the im- age increases. We present a Watermarking technique which requires no information for decoding in addition to the wa- termarked image. The watermark is multi-level with few bits embedded robustly at low levels and longer watermark sequences embedded less robustly at higher levels. This al- lows detection of ”tampered” and ”attacked” images by de- tection of existence of the watermark at low levels and dete- rioration of the watermark at high levels. In order to prevent interference of the watermarks at different levels various im- age representation spaces are used. The watermark is shown to be non visible and robust. 1. INTRODUCTION Digital Watermarking is a technique for marking and label- ing digital images. Methods have been developed to with- stand attacks on watermarked images including image pro- cessing, image compression and geometric transformations attacks. The various methods previously developed can be categorized according to the following characteristics: Representation Space - Watermarking embeds a code by changing the image. The embedding can be per- formed in the image represented in various domains. The domain chosen for image representation affects the robustness and capabilities of the watermark to withstand attacks. Watermarking in the spatial do- main [1, 2]has the advantage of maintaining locational information, thus geometrical attacks such as trans- lation and especially cropping are more likely to be overcome. Watermarking in the frequency domain [3, 4] does not preserve locational information but at- tends to the frequency content of the image enabling robustness under filtering such as blurring, high pass image enhancement, etc. Local frequency transforms of the image (such as block DCT and Wavelet Trans- forms) [5, 6] have the advantage of both locality and frequency extraction, but are highly sensitive to at- tacks involving translation or cropping of the image. Code Length - The watermark itself is a code varying in length from 1 bit (interpreted as: a watermark ex- ists or does not exist in the image), to a sequence of bits. The 1-bit watermarks (e.g. [1, 3]) are typically based on statistically evaluating the probability of the image containing a watermark. The watermark code containing a number of bits typically vary the values of the image representation in a sequence of changes that depend on a random generator and the bit values of the code (e.g. [2]. It is shown experimentally that 1-bit watermarks are far more robust under attacks than the watermarking techniques associated with a sequence of bits. Information Required for Decoding - A drawback of most watermarking techniques is the need for some additional information in order to retrieve the water- mark. In many cases, typically for the 1-bit water- marks, only the RNGS (random number generator seed) is required, and possibly the image size (e.g. [1]). For decoding of longer bit sequences, the wa- termarking techniques typically require, in addition to the RNGS, also the original image and/or the wa- termark code (e.g. [3] requires the original image the RGNS and the watermark code). The external information required for decoding puts a heavy re- striction on watermarking techniques, in that a list or database must be maintained to associate a given watermarked image with it’s watermark decoding pa- rameters (which might include the original image). Combining several watermarking techniques has already been suggested [7], however in this work, we exploit the advantages of various image representation domains to em- bed a multi-level watermark which is robust under cropping as well as filtering and compression attacks and requires no additional information for decoding. The watermarking