Article Preview
Top1. Introduction
Steganography is a security technique that utilizes digital files or network protocols to embed secret messages (Provos & Honeyman, 2003). Compared with traditional security technology, steganography has the advantage of concealment, which will make it undetectable for attackers. Accordingly, steganography can be applied to covert communication.
The research of steganography is mainly concentrated in images. Content-adaptive steganographic methods are the most secure schemes in recent years. Compare with traditional steganographic methods, content-adaptive steganographic methods can provide better security to resist the statistical detection. Filler, Judas and Fridrich (2010) developed a framework with Syndrome-Trellis Codes (STCs), which could be used for minimizing additive distortion between cover and stego images. There are many algorithms implemented by STCs, such as highly undetectable stego (HUGO) method (Bas, 2010), spatial-universal wavelet relative distortion (S-UNIWARD) method (Holub and Fridrich, 2013) et al. To enhance the security of covert communication, Sedighi, Cogranne and Fridrich (2016) proposed a method by using an estimated multivariate Gaussian cover image model to minimize the statistical detect ability. Content-adaptive image steganographic methods increase the difficulty of detection, but steganalysis technologies also make some progress in these years.
Rich-model based steganalysis is the modern methods for stego images detection. Fridrich and Kodovsky (2012) first design a rich-model based steganalysis method for images steganography. In their method, high dimensional features and ensemble classifier are employed to enhance the detection accuracy. Then Goljan, Fridrich and Cogranne (2014) designed an extension of the spatial rich model for color images. To detect the content-adaptive image steganographic methods, Denemark, Boroumand and Fridrich (2016) design some high order features by the knowledge of the selection channel. Luo et al. (2016) analysis the character of STCs and designed a steganalysis method for HUGO steganography. The method can not only detect the stego images but also extract the secret messages. Recently, Liu, Yang and Kang (2017) proposed a steganalysis method combines convolutional neural network with rich-models and ensemble classifiers. Experimental results show that the method has better performance than the state-of-the-art one. However, due to the structure and character differences between the parameters of image and speech, it is hard to directly employ the steganalysis methods on image to achieve effective detection for speech steganography.
In recent years, with the development of mobile network and smart phone, Voice over IP (VoIP) has become widely employed by mobile communication such as network telephone or instant message. Compared with other carriers for covert communications, VoIP has obvious advantages, for example, its large volume for embedding data could provide high covert bandwidth, and its instantaneity could provide real-time communication environment. Therefore, there are many works have been done for the steganography based on VoIP. As a standard of speech compression, AMR is widely employed by 3G, 4G systems or VoIP in speech services. Due to its great performance on speech compression, AMR is adopted as the file format for many communication applications such as intent message or speech recorder on smart phones. Therefore, the steganography of AMR speech codec has attracted extensive attention in recent years.