Distribution-Preserving Steganography Based on Text-to-Speech Generative Models

Steganography is the art and science of hiding secret messages in public communication so that presence cannot be detected. There are two distribution-preserving steganographic frameworks, one sampler-based other compression-based. The former requires a perfect sampler which yields data following same distribution, latter needs explicit distribution generative objects. However, these conditions too strict even unrealistic traditional environment, e.g., natural images hard to seize. Fortunately, models bring new vitality steganography, can serve as or provide media. Taking text-to-speech generation task an example, we propose steganography based on WaveGlow WaveRNN, corresponds categories. Steganalysis experiments theoretical analysis conducted demonstrate proposed methods preserve distribution.