Compressive sensing for data hiding

In many DRM schemes we embed a signature “watermark” in an image or audio file for copyright protection or content verification, to prevent the propagation of illegal media. In order to verify ownership we design detectors that can confidently establish the presence of a specific watermark.  We present a detection scheme that not only detects the presence of a watermark with 100% accuracy (under specified conditions), but does so with zero error in estimating it, and without knowledge of the image that is being watermarked or the watermark itself.

The central idea of our scheme stems from the idea of error-free reconstruction of sparse signals by ell1-decoding in the Compressive Sensing literature. Through ell1-decoding we can exactly recover a given input signal as long as it is only sparsely corrupted by error. As an extension it has also been shown that the error vector need not be sparse per se for perfect recovery; if it can be made compressible (its coefficients decay by a power law), then it will have a good sparse approximation and also can be recovered with small bounded error. As a result, the probability of error-free compressible signal recovery will be slightly compromised compared to sparse signal recovery. In ell1-decoding, the recovery algorithm does not require any information about the input signal; interpreting this in watermarking terms, the decoding procedure is blind to the host image.

Authors: Mona Sheikh, Richard Baraniuk

Publications: ICIP (2007)

Related research at Rice