Cryptospace Invertible Steganography with Conditional Generative Adversarial Networks
Deep neural networks have become the foundation of many modern intelligent systems. Recently, the author has explored adversarial learning for invertible steganography (ALIS) and demonstrated the potential of deep neural networks to reinvigorate an obsolete invertible steganographic method. With the worldwide popularisation of the Internet of things and cloud computing, invertible steganography can be recognised as a favourable way of facilitating data management and authentication due to the ability to embed information without causing permanent distortion. In light of growing concerns over cybersecurity, it is important to take a step forwards to investigate invertible steganography for encrypted data. Indeed, the multidisciplinary research in invertible steganography and cryptospace computing has received considerable attention. In this paper, we extend previous work and address the problem of cryptospace invertible steganography with deep neural networks. Specifically, we revisit a seminal work on cryptospace invertible steganography in which the problem of message decoding and image recovery is viewed as a type of binary classification. We formulate a general expression encompassing spatial, spectral, and structural analyses towards this particular classification problem and propose a novel discrimination function based on a recurrent conditional generative adversarial network (RCGAN) which predicts bit-planes with stacked neural networks in a top-down manner. Experimental results evaluate the performance of various discrimination functions and validate the superiority of neural-network-aided discrimination function in terms of classification accuracy.