DAIRstega: Dynamically Allocated Interval-Based Generative Linguistic Steganography with Roulette Wheel

Linguistic steganography (LS) tasks aim to generate steganographic text (stego) based on secret. Only authorized receivers can perceive and extract secrets, thereby protecting privacy. However, existing generative LS schemes often do not consider the conditional probability of tokens in the candidate pool, and allocate one or the same number of codings to all tokens. The tokens with lower probabilities are selected to embed secrets that will affect the quality of stegos. As a result, the stegos are easy to perceive and detect. This paper proposes the LS scheme based on dynamically allocated intervals, called DAIRstega. DAIRstega uses the idea of the roulette wheel and takes the conditional probabilities of tokens as the main basis for allocating the roulette area (i.e., the interval length). Thus, the token with a larger conditional probability is allocated more. The secret will be more likely to select the tokens with larger probabilities. In the allocation process, we design some functions between probability and allocated interval length. Based on the invisible characteristics of LS, we give three constraints that need to be met to design the function. To simplify the form, the expression of the allocation function is condensed. Furthermore, DAIRstega can receive additional instruction and controllably generate stegos. Rich experiments show that the proposed embedding way and DAIRstega perform superior to the existing ways and LS schemes, which shows strong perceptual, statistical, and semantic concealment and anti-steganalysis ability. This scheme can also generate high-quality longer stegos, improving the deficiencies in this task. The experiment also verified that DAIRstega can be used as a secure watermarking scheme, providing some ideas for its development.