To demonstrate the outcome of our research, we have selected representative of various watermarking methods with different ways of watermark insertion and representing a different class of zero-watermarking algorithms commonly used in medicine. As will be observable from the results below, all zero-watermarking algorithms except one provide similar results. Due to our long-term cooperation with Masaryk Memorial Cancer Institute, the algorithms examined were used on (anonymized) CT images of patient with tumor recommended by physicians from this institute.
During our testing of zero-watermarking methods, we conducted a lot of simulations and comparisons.
In the first phase, we focused on observing the behavior of individual zero-watermarking methods in case of modifying the secured image (see Fig. 3).
The result of this phase is shown as a graph in Fig. 5. Increasing the graphical modification of the secured image (from image numbers 2 to 40) decreases the PSNR (a metric of the similarity of the secured and modified images). Firstly, we gradually graphically removed a tumor (see Fig. 4); then, we made further gradual destructive changes in the medical image.
During this progress, we observed the behavior of the particular watermarking methods. In other words, we reduced the resemblance of the secured image and the modified image from which watermark is extracted (the PSNR), and we observed the similarity of the embedded and extracted watermark NCC (normalized cross correlation ). From the chart at Fig. 5, it is clear that the individual methods (except for zero CLBP ) can relatively correctly extract the watermark, even if the image is largely modified. In case of unauthorized modification of the secured image, it is, therefore, possible to prove its authorship by using suitable methods of zero watermarking. It is also possible to ensure integrity in case of modification because the NCC (normalized cross correlation) of modified and original images is never exactly equal to 1. For example, watermark extracted from modified image number 2 has NCC = 0.99978.
In the next phase, we compared characteristics of watermarking methods in the instance of exchanging watermarked images with neighboring CT images of the same patient. The graph in Fig. 6 shows that the similarity of the secured and changed images (from which the watermark is extracted) is low (PSNR < 40 dB; medical images are ordered from higher PSNR to lower). Nevertheless, individual methods of zero watermarking display relatively high NCC values between inserted and extracted watermarks. This means that the watermarks are quite similar.
These methods are therefore inappropriate to protect authorship when exchanging the watermarked image with neighboring images. This is because it is not possible to recognize if the medical image was only modified (and is owned by his author), or if it has been replaced with a different image of the same patient.
In the final stage, similarly to the previous phase, where we exchanged a former secured image with neighboring images, we exchanged a formerly secured image with an image of the same body part of a totally different patient. As in the previous phase, similarity of the secured and exchanged images is low (PSNR < 30 dB; medical images are ordered from higher PSNR to lower), the NCC values of the inserted and extracted watermarks were rather high, despite the low similarity of the secured and changed (attacked) images.
It is obvious from Fig. 7 that watermarking methods do not respond well even to this drastic form of image substitution and thus are not suitable for securing authorship.
The NCC itself is only an image similarity measure. If the NCC of inserted and extracted pair of watermarks is high, it still does not guaranty that the image the watermark was extracted from is a derivative of the original. It only expresses our belief or plausibility of being so. Currently, NCC above 0.7 is taken as the touchstone whether it is tested image suspected to be a derivative of the protected one. Proper determination of the threshold or possible two (soft and hard) thresholds is subject to further study. The Zero CLBP CliclePartition VC method comparing with others seems to be the much more sensitive in distinguishing modified picture from a completely different one.