A robust audio zero watermarking scheme using multi feature fingerprints and machine learning

一种基于多特征指纹和机器学习的稳健音频零水印方案

阅读:1

Abstract

The article presents a novel audio zero-watermarking technique that is free of creating any watermark data in the original audio. The strategy depends on the robust features of the audio in order to generate an audio fingerprint. Primarily, the audio is divided into non-overlapping frames, and time-based, frequency-based, and structural features are obtained out of each frame. The features are then weighted on the basis of its stability to construct a very robust composite feature representation. The adaptive thresholding scheme is used to apply a binary sequence to the composite representation and the watermark data is then appended to the binary sequence to produce the final audio fingerprint. Lastly, the features extracted along with the corresponding binary sequence that acts as a label of every frame are then fed to a Random Forest (RF) classifier. The feature vectors derived when extracting the attacked audio are fed to the trained Random Forest classifier in order to reconstruct the original binary sequence during extraction. This sequence is subsequently combined with the existing fingerprint and it becomes possible to reconstruct the watermark. The results of simulations indicate that the technique preserves the original quality of the audio, but provides good resistance to numerous types of attacks, such as additive noise, filtering, MP3 compression, re-sampling and re-quantization changes, as well as time and pitch rescaling.

特别声明

1、本页面内容包含部分的内容是基于公开信息的合理引用;引用内容仅为补充信息,不代表本站立场。

2、若认为本页面引用内容涉及侵权,请及时与本站联系,我们将第一时间处理。

3、其他媒体/个人如需使用本页面原创内容,需注明“来源:[生知库]”并获得授权;使用引用内容的,需自行联系原作者获得许可。

4、投稿及合作请联系:info@biocloudy.com。