A Distribution Detection Framework for Watermark Analysis

ABSTRACT Many watermarking algorithms have been proposed and studied for their robustness and other properties. But, there has been little e ort is analyzing these algorithms for their vulnerability against dete tion by an adversary. As a general philosophy for robust watermarking the host signal and the watermarked signal are well separated in a statisti al distan e sense. This very nature an be exploited by an adversary to easily dete t the watermark and perhaps remove it. We argue that the ability of a watermark to avoid dete tion by an adversary is a key fa tor that needs great attention. In this paper, we propose a method for watermark analysis based on a distributed dete tion framework that an be used by an adversary. The proposed framework an be used by the adversary to study signals for the presen e/absen e of watermarks. We hoose a parti ular spatial domain image watermarking algorithm and explain how the proposed framework an be applied to dete t the watermark with very little knowledge about the watermark insertion pro edure. The false alarm probability an be optimally traded-o with the probability of dete tion using the re eiver operating hara teristi fun tion.