A novel audio watermarking scheme using ensemble-based watermark detector and discrete wavelet transform

Most existing extraction techniques in audio watermarking use conventional techniques in which some sets of special rules based on reverse embedding rules are used for watermark extraction and have many weaknesses, like very low robustness to destructive attacks. To overcome this problem, the use of machine learning-based methods has increased in recent years in this field. The disadvantage of these methods is the high reliance on a unique classifier and lack of proper efficiency when achieving high capacity, which is a major challenge in audio watermarking. The main purpose of this paper is to present a method that covers the weak points of conventional methods and simple intelligent methods and improves system performance using a synergistic combination of discrete wavelet transform (DWT) and ensemble-intelligent extraction approach by proposed combination of trained machine learning classifiers. For the embedding operation in the proposed method, the DWT and the difference in energy levels obtained through DWT coefficients are used. In the extraction section, three methods are used in parallel: (a) the trained support vector machine (SVM) classifier with RBF kernel, (b) trained SVM classifier with quadratic kernel and (c) the trained K-nearest neighbor classifier; finally, the majority function is used to vote and make a final decision to create an intelligent-based watermark detector. A training set is required to train the classifiers, whose bit sequence is generated by a proposed 5-bit linear-feedback shift register. The results of various experiments indicate that this ensemble method has achieved the appropriate imperceptibility and high capacity, along with higher robustness compared to conventional techniques and individual learning classifiers.