A learning-based audio watermarking scheme using kernel Fisher discriminant analysis

摘要

A novel learning-based audio watermarking scheme using kernel Fisher discriminant analysis (KFDA) is proposed in this paper. Two techniques, down-sampling technique and energy relationship modulation technique, are developed in order to guarantee good fidelity of the watermarked audio signal. At the same time, local energy relationship between audio sub-frames is hid in the watermarked audio signal with watermark embedding. Moreover, a learning-based watermark detector using the KFDA is exploited and it extracts the watermark by learning the local energy relationship hid in the watermarked audio signal. Due to powerful non-linear learning ability and good generalization ability of the KFDA, the learning-based watermark detector can exhibit high robustness against common audio signal processing or attacks compared with other audio watermarking methods. In addition, it also has simple implementation and lower computation complexity.