Digital Audio Watermarking Techniques Based On Machine Learning Method

With rapid development of Internet and digital multimedia technique, we can conveniently and quickly obtain, edit, reproduce and spread more and more digital audio. It is beneficial to our daily life but brings much serious security problem, such as ownership protection of digital audio. Digital audio watermarking technique, as a efficient tools for protecting the intellectual property rights of digital audio data, has very important application prospect. In the recent years, digital audio watermarking technique is developed fast. However, some key issues are still open. This dissertation pays more attention to robust audio watermarking techniques based on machine learning and their some key issues, and present three novel audio watermarking approaches based on machine learning. The main contributions are as follows:(1) Aiming at the problem that some machine learning methods have complex training algorithms and long training times, such as artificial neural network, support vector machine, etc., an audio watermarking approach using kernel discriminant analysis is proposed. On basis of down-sampling technique, an energy modulation technique is developed, including two parts, one is modulation of energy relationship, and another is proportion modulation method of DWT coefficients. The modulation technique can bring the benefit that it can guarantee the imperceptibility of watermarking system (i.e., auditory quality of watermarked audio) and hide some energy relationship into audio signal. The watermark detector based on kernel discriminant analysis can extract watermark signal by learning the energy relationship hid in audio signal. Due to powerful (nonlinear) learning ability and good generalization ability of kernel discriminant analysis, the watermark detector has high robustness against attacks, and its training algorithm is simple and fast.(2) By regarding the tradeoff between imperceptibility and robustness against common signal processing and desychronization attacks simultaneously as optimized objectives, an audio watermarking scheme in multiwavelet domain is proposed, which can efficiently resist common signal processing and desychronization attacks. Firstly, the mean quantization technique to embed synchronization code in temporal domain is used, which is used to determine embedding positions of watermark signal in order to resist the damage of desychronization attacks for the embedding positions. Secondly, an energy quantization modulation method is developed to embed the watermark. The modulation method regards local energy index of low-frequency multiwavelet coefficients as the modulation object, and deduces a coefficient modification formula of corresponding low-frequency multiwavelet coefficients according to algebra formula. Finally, an optimization procedure of audio watermarking developed by using particle swarm optimization (PSO) automatically determines optimal watermarking parameters of watermarking algorithm, which can optimally balance imperceptibility and robustness against both common signal processing and desychronization attacks.(3) The optimal watermarking problem is a multi-objective optimization problem essentially. However, previous genetic algorithm (GA)-based watermarking methods or schemes convert the optimal watermarking problem into a single-objective optimization problem by using a weighted sum form. Aiming at multi-objective essence of optimal watermarking problem, a multi-objective optimization framework to enhance performance of audio watermarking algorithm is proposed. On basis of non-dominated sorting genetic algorithm II (NSGA-II), a multi-objective genetic algorithm with variable-length mechanism and corresponding audio watermarking optimization procedure is developed, including variable-length chromosome, initialization mechanism, hybrid crossover and mutation operation. The conventional watermarking algorithms are easily incorporated into the multi-objective optimization framework, which can automatically determine their optimal watermarking parameters and search most suitable embedding positions. In addition, multiple solution property of Pareto optimal set can provide more flexibility for designing watermarking system.