Wavelet Based Multi-bit Fingerprinting Against Geometric Distortions

This paper presents a new image fingerprinting scheme which embeds a multi-bits fingerprinting code and is robust against the geometric attack such as rotation, scaling and translation. We construct a 64 bits fingerprinting code and embed into wavelet subband of 512x512 images repeatedly. In order to restore an image from geometric distortion a noise reduction filter is performed and a rectilinear tiling pattern is used as a template. Results of experimental studies show that our method is robust against geometric distortions and JPEG compression. Introduction Digital fingerprinting is one possible application of data embedding techniques, whereby some unique information, such as a serial number or a user ID assigned by the vendor to a given user/purchaser, is embedded into the multimedia content using watermarking techniques. One powerful class of attacks that adversaries may employ against watermarks and the corresponding fingerprints is collusion [1,2], whereby a coalition of users combines their different marked copies of the same multimedia contents in an attempt to attenuate/remove the trace of any original fingerprint. The fingerprint must, therefore, survive both standard distortions (such as compression, filtering and geometric distortion) and collusion attacks by users intending to destroy it. Several methods have been proposed in the literature to embed fingerprints (watermarks) into different media and different domains [1,2]. But, these methods are not enough to present the variety of customer information as a fingerprint. To identify lots of customer’s multi-bits embedding scheme is required. Moreover, RST distortions are also a big problem to be solved in the image watermarking and fingerprinting area [3]. In this paper, we propose a new image fingerprinting scheme which embeds 64-bits customer ID into the discrete wavelet transform (DWT) domain and extract this ID from the geometrically distorted image using ACF(Auto Correlation Function). The paper is organized as follows. How to embedding and extracting fingerprint information are presented in Section 2. The simulation results and conclusions are given in Section 3 and 4, respectively. Proposed Fingerprinting Scheme Fingerprint Embedding. In this section, we describe the way of 64-bits fingerprint construction and embedding procedure. Assume that a 64-bits message which can be separated as 8 symbols whose length is 8 bits is given. We suppose that each symbol has Alphabet capital and small letters and numeric numbers from 0 to 9, S={s1, s2, s3,...,sN} {a,...,z,A,...,Z,0,...,9}. Then, we generate random sequences for every symbol from a secret key which can be represented in ri {-1, +1}. Therefore, the total number 3 * Corresponding author Key Engineering Materials Vols. 321-323 (2006) pp 1301-1305 online at http://www.scientific.net © (2006) Trans Tech Publications, Switzerland Online available since 2006/Oct/15 All rights reserved. No part of contents of this paper may be reproduced or transmitted in any form or by any means without the written permission of the publisher: Trans Tech Publications Ltd, Switzerland, www.ttp.net. (ID: 130.203.133.34-16/04/08,09:20:56) of ri is (26×2+10)*8+1 = 497. Here, one random sequence is a sync message for the translation restoration. Finally, in order to produce fingerprint signal, W, the 9 random sequences corresponding to fingerprint symbols are merged together and the sign of the merged sequence is taken. In order to be robust against JPEG compression the periodic fingerprint patterns are embedded into the detail subbands of wavelet decomposition level 2. The overall process of the proposed fingerprint embedding is shown in Figure 1. To extract fingerprint correctly from the corrupted image, it is important to restore the image from geometrical distortions. In the proposed system a rectilinear tiling pattern is used to do this. To construct the latticed template random sequence is embedded repeatedly as shape of unit block. Embedding unit blocks are used as a template to restore the captured image. In this paper the size of unit block is 32x32. For instance, if the image size is 512x512, then unit block is embedded 16 times repeatedly in the HL2, LH2, and HH2 sub-bands. Fig. 1 Fingerprint embedding process. HVS(Human Visual System) is a weighted function to make the image robust to various kinds of attacks and to improve imperceptibility. The basic idea of our HVS is that fingerprint is embedded strongly into the less recognizable regions of the image. To do this, we separate the image into three regions; flat, strong edge, and texture region according to edge detection value. If the edge detection value, Edge(i,j) is smaller than 2, we set position (i,j) to flat. On the contrary if the edge detection value is bigger than 2, we set position (i,j) to texture area. Especially if the edge detection value is bigger than threshold T, we set position (i,j) to strong edge area. T is defined as follows: ) ( * 2 ) ( I StD I Aver T Edge Edge + = Aver(i,j), StD(i,j), and Edge(i,j) are local average, standard deviation, and edge detection value on x(i,j) respectively. For edge detection Prewitt operator is used. AverEdge(I) and StDEdge(I) are average and standard deviation of edge detection values. HVS function is as follows:    = − = = ] 255 ... 0 [ , 3 / )) 25 / tanh( 2 ( ) ( , )) , ( ( * ) , ( , ) , ( i i i WF where Otherwise j i Avg WF j i StD area edge Strong Flat j i α λ α is minimum embedding strength and set to 3. WF(*) is a weighted function for dark and bright area. So, fingerprint is embedded strongly by this function because these areas are less sensitive than normal area. Fingerprint Extracting. In this paper general correlation detector is used to extract fingerprint. But, pre-processing is needed before extracting because the fingerprinted image includes various kinds of distortions. Although we embed the fingerprint into the DWT domain, we can extract the periodicity of the fingerprint like the spatial domain method by using a high-pass filter or a noise removal filter. In our method, the periodic signal is extracted by using a Wiener filter, and the average of the local variances is used as the noise variance of the Wiener filter. 1302 Advanced Nondestructive Evaluation I