SmartSteganogaphy: Light-weight generative audio steganography model for smart embedding application

Abstract

Massive data is transferred in the Internet of Thing (IoT) every single second. Protecting the security of these data is a crucial task. Steganography is a collection of techniques for concealing the existence of information by embedding it within irrelevant carriers, which could protect security and privacy of the data. Distinct from the cryptography, steganography put emphasis on hiding the existence of the secret. However, the majority of research mainly work on complex algorithm. In this paper, we proposed an audio steganography algorithm which automatically generated from adversarial training. The embedding model is light-weight which could use as machine learning tools in smart device. Besides, the existing audio steganography methods mainly depend on human handcraft, while the proposed method could obtain from meachine learning. The embedding model consists of three neural networks: encoder which embeds the secret message in the carrier, decoder which extracts the message, and discriminator which determines the carriers containing secret messages. The system is trained with different training settings on two datasets. Competed the majority of audio steganographic schemes, the proposed scheme could produce high fidelity steganographic audio. Besides, the extensive experiments demonstrate the robustness and security of our algorithm.

Introduction

The Internet of Things is a network of connected software, smart device, and many items which facilitate the data exchange (Ashtonet al., 2009). The main purpose of IoT is to provide the security and reliability of data transmission thus control Things. Although IoT provide the convenience of life, the security needs attention. Massive data is transferred in the IoT every single second which vulnerable to attack. Personal information could be hacked from the system without protection and therefore protecting the security of these data in IoT is a crucial task (Zhao and Ge, 2013; Kadhim et al., 2020). The cryptography techniques are wildly used in many identification and authentication applications in IoT to ensure the security. However, massive surveillance operations have shown that the mere existence of meta-data communication could lead to privacy leakages even if the content is unknown. To further protect the data security, we introduce steganography techniques in IoT which focus on hiding the existence of the information in proposed work.

Steganography is the science to conceal secret messages in the irrelevant carriers through slightly modifying the values which hard to detect by human perception. Similar to cryptography, the steganography provides methods for secret communication. Distinct from the cryptography method which focuses on the authenticity and integrity of the messages, steganography aims to hide the existence of the secret. Steganography hides the encrypted data that anyone would suspect the existence of the secret through special algorithms to provide additional security (Hurrah et al., 2019). In this work, we proposed a novel framework that automatically generated light-weight audio steganography algorithms. Audio steganography algorithms embed secrets into audio contents based on the elusion of human auditory system. Audio steganography could be used in the watermark, copyright protection and secret transmission and many other applications. Existing audio steganography methods could be divided into three categories: temporal domain methods which encoder messages in Least Significant Bit (LSB) of individual sound samples with the equivalent secret message binary sequence (Balgurgi and Jagtap, 2013; Bender et al., 1996), transform domain methods which use the masking effect of human auditory system and make the weak frequencies near the strong resonance frequencies inaudible (Alwahbani and Elshoush, 2016), and wavelet domain methods which use discrete wavelet transform and hide the information in discrete wavelet coefficients adaptively (Sheikhan et al., 2010).

Steganography and steganalysis are import topics in information hiding. In the early years, steganography is relatively simple and hence the dimension of feature in steganalysis was low and light computation. With the study of content adaptive steganographic schemes which usually employ distortion functions and syndrome-trellis coding technique to embed secret messages in relatively secure locations of the carrier. Classic methods mainly depend on human handcraft, and most of the techniques take disadvantage of low capacity, lack of security and fixed generalization. With the development of deep learning, many in-depth models have used in steganography and steganalysis (Zou et al., 2019; Zhu et al., 2018; Boroumand et al., 2018). Generative adversarial networks (GANs) have proved to be competitive generative models on synthesis tasks in recent researches (Goodfellow et al., 2014; Isola et al., 2017; Radford et al., 1511). GANs have also applied to speech generation nor enhancement task and generate signal (Yang et al., 1703; Pascual et al., 1703; Hsu et al., 1704). In this work, we introduce the idea of adverasrial training to automatically generated audio steganography algorithm.

Distinct from the most audio synthesis task, the target of audio steganography is embedding with less distortion. We introduce the idea of zero-sum game theory and propose the embedding model which consists of three neural networks: encoder which embeds the secret message in the carrier, decoder which extracts the message, and discriminator which determines the carriers containing secret messages. The embedding model aims to conceal the secret audio within carrier audio. Thus, the task of the training is a binary classification problem. The encoder takes in carrier audio and secret message then produces steganographic audio, while the discriminator tries to learn the weakness of generator, resulting in the ability to determine whether the audio contain secrets. All the networks are simultaneously trained on the datasets until the encoder could produce high fidelity steganographic audio. We show that our scheme could successfully work in practice, the encoder could produce steganographic audio which contains secret audio, the decoder could decode the steganographic audio to a carrier and the less distortion secret audio, which indicate heard more than heard. As far as we are concerned, GAN has not yet been applied to audio information hiding task, so this is the first approach to use the adversarial framework to generate steganographic audio. Distinct from most meachine learning model, the proposed model is less than 5 MB which could used in many smart device in IoT.

The paper is organized as follows. In Section 2, we present a review of steganography and steganalysis. The structure of our model and methodology is described in Section 3, which is followed by the experimental results and analysis in Section 4. Finally, the concluding remarks are drawn in Section 5.