Blind Audio Watermarking Algorithm Based On Machine Learning

The digital media has become a main way for information communication along with the rapid development of digital technology and computer networks. As a useful tool for the copyright protection and judgment, digital watermarking technique has gained more and more concerns in many ways since it appeared in 1993, meanwhile, protection of digital multimedia information has become an increasingly important issue. Especially audio digital watermarking technique has been one of the research hotspots in recent years. Traditional information security system can only safeguard information transmitting process, but which can't control the decoded media data. So it can't prevent the illegal copy of the pirate. As a novel way to solve these problems, digital watermarking technology begins to be popularly researched and used. By embedding some secret watermark information in the host multimedia signals, it provides solutions to copyright protection and content verification.With the digital audio watermarking technique we can embed secret information into digital audio signal, so as to arrive at the purpose of copyright protection and judgment. The chose of the work field and the embed method becomes more necessary. Wavelet transform, as a new powerful tool of time-frequency analysis, provides several good characters that make it appropriate to audio signal watermarking. So, a novel machine learning based digital blind audio watermarking scheme in the wavelet domain is proposed in this paper, which could find a reasonable balance between the robustness and inaudibility of the watermark, then have the better robustness and inaudibility. And the proposed watermarking method which doesn't require the use of the original audio signal for watermark extraction also can provide a good copyright protection scheme.The main contribution of this paper is as follows:(1) This paper focuses mainly on a novel support vector regression (SVR) based digital audio watermarking scheme in the wavelet domain which using subsampling. The audio signal is subsampled firstly and all the sub-audios are decomposed into the wavelet domain respectively. Then the watermark information is embedded into the low-frequency region of random one sub-audio. With the high correlation among the sub-audios, accordingly, the distributing rule of different sub-audios in the wavelet domain is similar to each other, SVR can be used to learn the characteristics of them. Using the information of unmodified template positions in the low-frequency region of the wavelet domain, the SVR can be trained well. Thanks to the good learning ability of SVR, the watermark can be correctly extracted under several different attacks, and the proposed watermarking method which doesn't require the use of the original audio signal.(2) Aiming at the problem which is the robustness and inaudibility of the digital audio watermark is limited for each other, so an optimization watermarking method based on Genetic Algorithm (Genetic Algorithms, GA) which to compute the best energy quickly is proposed in this paper. Because the genetic algorithms could search the optimization of objective function randomly, which via the evolution of groups, our work is based on the embedding/detecting way of "DWT-Based Audio Watermarking Using Support Vector Regression and Subsampling". And this method provides a preliminary discussion on the application of GA, which searching the individual with the highest fitness of attack-resistance in the aggregate of the optimal embedding strength. As an optimization solution of obtaining the optimal embedding strength, it could achieve an adaptive policy, and find a reasonable balance between the robustness and inaudibility of the watermark, just for the better robustness and inaudibility in the algorithm.The experimental results show the two practical algorithms can preserve inaudibility and they are robust enough to against the different signal processing operations. In addition, the watermarking information can be embedded into the original audio signal adaptively, which also doesn't require the use of the original audio signal for detection. Furthermore, it can resist the different attacks, such as lossy compression (MP3), filtering, resampling and requantizing, etc. So it can be used for copyright protection of digital audio productions.