The Syndrome-Trellis Sampler for Generative Steganography

We adapt the Syndrome-Trellis Code algorithm to generative steganography, giving a method for sampling from a specified distribution subject to linear constraints. This allows the use of syndrome codes, popular in cover-modification methods, for cover-generation steganography. The SyndromeTrellis Sampler works directly on independent and Markov-chain distributions, and can be plugged into an existing STC-based method to extend it to Gibbs fields that can be decomposed into conditionally-independent sublattices. We give some experiments to show that the method is correct, and to quantify how the payload condition forces the sampled distribution away from the target. The results show that the secrecy of the parity-check matrix of the syndrome code is important. We also show how to exploit sparsity in the conditional cover distribution, in a simple example from linguistic steganography.