Security evaluation of a perceptual image hashing scheme based on virtual watermark detection

This paper evaluates the security of a recently proposed perceptual image hashing scheme based on virtual watermark detection. Under the known-hash attack where the attacker has access to several image/hash vector pairs, we show that the task of estimating the virtual watermark sequences serving as the secret key can be formulated as a simple convex optimization problem, and hence, can be solved efficiently. More specifically, we demonstrate that satisfactory level of estimation accuracy of a watermark sequence of length m could be achieved from approximately 2 · m image/hash vector pairs on average. Experimental results using artificial data and real image data are also provided to verify the effectiveness of our proposed attack approaches.