基于多层激励量子神经网络的鲁棒音频水印算法

Improved quantum neural networks based audio watermarking algorithm

设计了一种基于多层激励函数量子神经网络的音频水印算法。将水印信号嵌入载体语音的小波低频系数中,再训练量子神经网络建立水印嵌入前后低频小波系数间的联系以便在接收端恢复水印。同时,区别于传统的归一化方法,将小波低频系数规范到同一数量级,避免了恢复水印时小波低频系数的差错传播,提高了算法的鲁棒性。实验结果表明,设计的水印算法对加噪、滤波、重采样和再量化等攻击具有较强的鲁棒性,提取正确率相比BP网络水印算法平均提高1.8%。

A novel MAF-QNN （multilevel activation function-quantum neural networks） audio watermarking algorithm was proposed in this paper.Firstly,the watermark was embedded into the low frequency wavelet coefficients of the carrier speech.Then the QNN was trained to establish the contact between the low-frequency wavelet coefficients before and after the watermark embedded.By this way,the watermarking can be recovered at the receiving end.Distinctive from the conventional normalization method,the low frequency wavelet coefficients are normalized to the same order of magnitude,thus the error propagation of the low frequency wavelet coefficient can be avoided while the watermark recovering,and the robustness of the algorithm is improved.Experimental results show that,the watermarking algorithm designed in this paper is robust against some different attacks effectively,such as adding noise,filtering,re-sampling and re-quantization attacks.Furthermore,the correct rate of watermarking extraction is increased by an average of 1.8％ compared with the BP neural network watermark.