Digital watermark extraction using support vector machine with principal component analysis based feature reduction

• Efficient especially in case of JPEG compression attack with low quality factor.
• Also provides more robustness against various signal processing operations.
• Randomization using different secret keys helps the system to be more secure.
• It uses simpler features set for training and testing the SVM.
• Shows noteworthy comparisons with currently existing techniques.

This paper proposes a new approach for watermark extraction using support vector machine (SVM) with principal component analysis (PCA) based feature reduction. In this method, the original cover image is decomposed up to three level using lifting wavelet transform (LWT), and lowpass subband is selected for data hiding purpose. The lowpass subband is divided into small blocks, and a binary watermark is embedded into the original cover image by quantizing the two maximum coefficients of the block. In order to extract watermark bits with maximum correlation, SVM based binary classification approach is incorporated. The training and testing patterns are constructed by employing a reduced set of features along with block coefficients. Firstly, different features are obtained by evaluating the statistical parameters of each block coefficients, and then PCA is utilized to reduce this feature set. As far as security is concerned, randomization of coefficients, blocks, and watermark bits enhances the security of system. Furthermore, energy compaction property of LWT increases the robustness in comparison to conventional wavelet transform. A comparison of the proposed method with some of the recent techniques shows remarkable improvement in terms of robustness and security of the watermark.