Flaw in SVD-based Watermarking

Abstract

In the recent years, many papers have been published on the use of Singular Value Decomposition (SVD) for watermarking because of its robust nature. The singular values that are produced are very stable and vary very little under attacks. This introduces an ideal medium for which a watermark is embedded for robust watermarking. Nevertheless a severe flaw has been discovered by Zhang and Li in [3], which is based on an image watermarking technique proposed by Liu and Tan in [2]. In this paper the discovered flaw is explored and tested for a SVD-based audio watermarking technique proposed by Ozer et al. in [1]. Ozer proposed a technique in which the audio signal is first transformed into matrix form using the Short-Time Fourier Transform (STFT). The SVD is then used to decompose the STFT matrix in order to produce its singular values. As the watermark is embedded, two matrices are formed as a byproduct of the watermark and sent to the detector along with the watermarked signal. The flaw arises in the detection stage of this technique. Zhang argues that the detection of the watermark is highly dependent on the passed matrices and this introduces a high rate of false positive detection. Experiments show that the detection stage depends primarily on the passed information and depends very little on the watermarked signal. Therefore by altering the watermarked signal with various attacks gives a false sense of robustness and can easily be seen as an extremely robust system.