科研之友
微信
新浪微博
Facebook
分享链接
摘要
Audio steganography, which aims at hiding the secret information (i.e., text, image, audio, video, etc) into the audio carrier to not only ensure the security of the secret information itself, but also guarantee the transmission security, has become one of the hot topics in the field of information hiding. In recent years, the deep learning based audio steganalysis exhibits high detection efficiency by fully mining the deep features of steganography, which reduces the steganography security and brings new challenges to steganography. Fortunately, the rapid development of generative adversarial networks (GAN) provides a new solution to audio steganography. However, the existing GAN based audio steganography approaches cannot achieve a balance between hiding capacity, imperceptibility, and anti-detection ability, which is difficult to meet the needs of actual applications. In this work, an adaptive audio steganography approach, named BNSNGAN, is proposed on the basis of the optimized spectral normalization GAN by combining batch normalization(BN) with spectrum normalization(SN) in the network structure unit. Specifically, a steganographic encoder is first designed, in which the zero-padding in the time domain is used to preprocesses the secret audio to embed the secret audio with arbitrary length, thus improving the imperceptibility. Secondly, a steganographic extractor with a parallel structure is designed, in which different convolution cores are adopted for deconvolution and improving the accuracy of secret information extraction. Then, a steganalyzer with a cross entropy based loss function is designed to improve the anti-detection ability of audio steganography. Finally, comparative experiments demonstrate that according to the mutual learning of encoding, extractor, and steganography analyzer, the proposed BNSNGAN is able to embed the secret audio with arbitrary length, has a high rate of secret information extraction, and achieves a good balance between hiding capacity, imperceptibility, and anti-detection ability. ? 2022, Science Press. All right reserved.
- 单位
