Testing Results

We will now discuss and show graphically results of subjecting the smoking lady image to a variety of non-malicious and malicious attacks. The method of attack will be described and possible reasons for why one algorithm achieved stronger correlation over the other will try to be discovered.

Testing Summary

Testing the Cox and Xie watermarking algorithms weaknesses and strengths were uncovered. Cox was found to be stronger of the two algorithms tested. Both Cox's and Xie's watermark was vulnerable against cropping and rotation. Cox's algorithm faired better than Xie's against the lower bit level attacks such as Linear Geometric Transformation, row and column removal and NxN median filtering. Cox's algorithm was found to be more robust, one reason being that it stores the watermark information in significant regions of the DCT were as Xie's manipulates a sliding window of bits in the DWT that can easily be changed. Another weakness of Xie's algorithm found was that a picture only has to be cropped 1% to completely destroy watermark recovery process. This is true because cropping the Xie watermarked image slightly destroys the watermark reconstruction point that is needed to extract the watermark correctly. Cox's watermark was affected by cropping when most of the watermark had been cropped out. Even though Cox's algorithm was one of the first robust watermarking algorithms developed it was determined to be superior to Xie's algorithm.

Watermarking Strengths and Original Correlations

The un-attacked watermarked images correlations differed between the Xie and Cox algorithm. Parameters for the algorithms were adjusted to achieve the best-correlated and unambiguous watermark possible. The embedding strength of Xie was set to 0.7 and the watermark length was set to 800. When embedding strength was set to above 0.7, say 0.8, the watermarked image was very blotchy and the watermark was definitely noticeable. The un-attacked extracted watermark had a correlation of 0.899999. The embedding strength of the Cox watermark was set at 0.5 and length of the watermark was 800. When trying to increase length of watermark or embedding strength led to a blotchier watermarked image. The un-attacked extracted watermark had a correlation of 0.976267.

NxN Median Filtering
Generally the purpose of image filtering is to perceptually enhance features, like edges, or to suppress uninteresting features, such as blurring. Filtering is usually performed by a moving a window of NxN pixels, which is centered over each pixel of the input image. We applied NxN median filtering to the smoking lady image using Ifranview.
Median filtering returns the median pixel value in the moving window. The following results were obtained:

Figure #1: Xie NxN Median Filtering

Figure #2: Cox NxN Median Filtering

Figure #3: Unfiltered and 11x11 Median Filtered Images

As can be seen from the above results both algorithms are very robust against NxN median filtering with N <= 11. As N increases the median filter blurs the image considerably as can be noted in the above figure. The Xie's correlation varies much more than Cox's correlation with respect to N. This is most likely due to the Xie watermark being embedded in the DWT and the median filter manipulates the pixels directly therefore altering the watermark.