Sensitivity attack to watermarking schemes with nonparametric detection boundaries

The signals from which sensitivity attacks are launched directly determine the computational errors in attacks. With mathematical models on the accuracy of watermark estimations, it is shown more accurate to estimate watermarks from high-detection-value signals with fixed absolute errors. For watermarking schemes with nonparametric detection boundaries, sensitivity attacks directly launched from object signals can eliminate the negative effect of disordered boundaries, and the improved attack is proved effective in removing watermarks there.