Three-State Information Hiding: Provably Secure Asymmetric Steganography

The rise of language models has provided a fertile ground for the application of steganography. Due to their qualified output, steganographic texts become similar to human and have attracted most of the steganography researchers' attention. However, running a language model requires a strong computation platform. It limits the applicable scenario of steganography, since those electronic devices controlled by the decoder may not even equipped with a GPU. Traditional provably secure steganography methods cannot be applied to this low-resource scenario. Therefore, we aim at design a novel steganography framework that is practical in a low-resource scheme. We start from the rigorous probability analysis with the help of hypothesis testing techniques to construct an theoretical framework. Then we prove the security and robostness of our framework and point out its optimization goal. We test our theoretical framework in some famous LLMs and the results have proved its usability. There are still some practical problems and this gives the direction of future work. We hope that this work will expand the practical scope of steganography and create a new branch of steganography.