Font Size: a A A

A Research On Authentication Audio Watermark Algorithm Based On Deep Learning

Posted on:2020-03-07Degree:MasterType:Thesis
Country:ChinaCandidate:B JiangFull Text:PDF
GTID:2518306518963149Subject:Computer technology
Abstract/Summary:PDF Full Text Request
Digital watermark is an effective way to protect the information security,and considering audio is often used type of media,there's great potential on the research of authentication audio watermark.Current watermark is lack of security since it's static,and the study of audio watermark based on deep learning still waits for exploration.The authentication audio watermark algorithm based on deep learning is mainly studied in this paper.For watermark generation,this paper explores the feasibility of obtaining identity feature from audio as watermark by adopting the idea of voice profiling,and proposes an identity watermark generation model based on generative adversarial network.This model visualizes human face images by utilizing such identity feature,and develops antagonistic and constrained training under the designed discriminator and classifier in order to ensure the reliability of watermark as identity feature.For watermark embedding and extraction,differences in traditional mathematic methods,a watermark embedding and extraction model based on autoencoder by adopting deep learning is proposed.Embedding network and extraction network,included in this model,are updated and optimized through a fine designed loss function involving audio and watermark.The authentication watermark is adaptively embedded in audio through embedding network,following by the authentication of watermark which reconstructed from the extraction network.Based on the VoxCeleb English dataset and the self-collected Chinese multimodal dataset,the results of experiments,including authentication audio watermark generation,embedding and extraction,demonstrate the effectiveness of proposed authentication audio watermark algorithm based on deep learning.The authentication watermark obtained from generation model can achieve a relatively high within-class similarity and between-class dissimilarity.And the embedding and extraction model can complete the processes of embedding and extraction adaptively without sacrificing the imperceptibility and robustness.
Keywords/Search Tags:Audio Watermark, Identity Authentication, Dynamic Watermark, Adaptive Embedding, Deep Learning
PDF Full Text Request
Related items