Digital watermarking is an enabling technology to prove ownership on copyrighted material, detect originators of illegally made copies, monitor the usage of copyrighted multimedia data, and analyze the spread spectrum of data over networks and servers. Most watermarking methods for images and video can be viewed as a communications problem in which the watermark must be transmitted and received through a watermark channel. This channel includes distortions resulting from attacks and interference from the original digital data. It is well accepted that an effective watermarking scheme may be described as the secure, imperceptible, robust communication of information by direct embedding in, and retrieval from, digital data. For verifying the security and robustness of watermarking algorithms, specific attacks have to be applied to test them. By using a theoretical approach based on random processes, signal processes, and communication theory, we propose a stochastic formulation of a new watermarking attack using a blind source separation-based concept. The proposed attack considers the watermarking channel as a "black-box". A host image was passed through the "black-box", which includes the watermarking embedding process, and then the watermarked image was produced The watermarked image is viewed as linear mixtures of unknown source signals, and we attempt to recover sources from their linear mixtures without resorting to any prior knowledge by using blind source separation theory.