Watermarking technique.
Watermarking technique
Initially, the technique of marking images with digital «watermarks» (and later digital audio signals and video data) was developed to protect copyrights). This article describes an unconventional way of using «watermarks».
In this article, we propose to consider the technique of «watermarks» for detecting distortions in digital images. This technique allows us to identify side (distorting) changes in properties that have appeared as a result of image processing by comparing correlated values from different parts of the image. The implementation of this technique requires a small amount of memory and small computing power, which makes it possible to use it for hardware implementation in digital cameras.
The technique involves dividing an image into blocks and marking these blocks with watermarks that depend on a secret key (camera ID) and the length of the image. This watermarking method is based on the use of wideband signals. To achieve continuous dependence on the image, we propose a special process of extracting individual bits from each block taking into account a threshold value generated based on key-dependent random smooth patterns. These bits are then used to initialize the PRNG and synthesize the wideband signal.
1. Introduction
Powerful publicly available graphic packages, such as Adobe Photoshop or Paintshop Pro, have various tools in their arsenal that can seriously change real images. This can be a combination of different parts of one or more images, and so high-quality that the boundaries are simply invisible. Such modification (montage) of an image can in some cases be detected when examining the noise of different parts of the image by comparing the histogram of non-intersecting image blocks or by searching for a violation of the continuity of areas. However, experienced forgers do the montage in such a way that it is almost impossible to detect the forgery.
The possibility of image manipulation is one of the reasons why they are never accepted as evidence. On the other hand, in some fields (especially for the military) it is very important to know whether the received digital image has been forged or not.
To effectively detect image forgery, a technique called «watermarking» can be used, with the help of which small blocks of the image are marked. «Watermarks» will depend on a secret key, which is then used when detecting modification of the image.
One of the first techniques used to detect image corruption (modification) was the technique based on the embedding of checksums in the least significant bit (LSB). Welton [1] proposed a technique that uses a key-dependent pseudo-random sequence that «walks» over the image. The checksum is constructed from the seven most significant bits and is inserted into the LSB of selected pixels. The checksum is «walking» in order to prevent modification of exchanging groups of pixels with the same checksum. However, although checksums can provide a very high probability of detecting changes, they cannot distinguish between brightness adjustment and modification that changes a face in the image. Increasing the grayscale for all pixels gives more significant changes, even if the image remains unchanged and usable.
Van Schindel et al. [2] modified the LSB of pixels by adding a long t-sequence to the pixel rows, the phase of which carries the watermark information. A simple cross-correlation is used as a test for the presence of a watermark. This method will provide a low level of security for any LSB technique and will not be robust to image processing operations with low-frequency characteristics.
Wolfgang and Delne [3] extended the work of Van Schyndel and improved the localization and robustness properties. They used m-sequences in the range (-1,1) arranged in 8×8 pixel blocks and appended them to the corresponding image blocks. Their technique is moderately robust to linear and nonlinear filtering and small amounts of added noise. However, since the watermark is inserted in the LSB plane, it can be easily removed.
[4] proposed techniques based on the use of masking in the spatial and frequency domains. Such «watermarks» are guaranteed to be invisible, but nevertheless allow detecting errors of more than half of the maximum permissible changes in each pixel or frequency, depending on the masking technique used. The image is divided into blocks, each of which contains a secret random signature modulated by the masking value of this block. In this case, the error estimate for small distortions is quite accurate.
However, it is unclear whether this technique can provide the necessary information for images in which distortions are visible to the eye.
In this case, it is better to use a more robust watermarking scheme applied to large blocks. The watermarks in the described method [4] are weakly image-dependent.
The secret signature should not depend on the image — it is modulated by the masking values of each block. But these masking values are available and can be easily calculated. Marking a large number of images with one secret key will not be secure, so this technique cannot be used in digital cameras.
In this paper, we describe a technique that uses robust watermarking for large blocks (64×64 pixels). To prevent unauthorized removal or distortion, the watermark should depend on a secret key S (ID), the number of blocks B, and the block content. The content of each block is represented with M bits extracted from the block using a random smooth template and a certain threshold. The result is an M-tuple (an ordered set) of such blocks, which allows successfully synthesizing a wideband signal of a marked distorted image. The wideband signal for each block is generated by adding a pseudo-random M-sequence uniformly distributed in (-1, 1). Moreover, each sequence depends on the secret key, the block number, and the bit extracted from the block. If k of the M bits are extracted incorrectly due to image distortion, the wideband signal will have a high correlation with the image as long as k