Provably Secure Linguistic Steganography via Range Coding Achieves Optimal Capacity

arXiv·low signal

Directly applies range coding to language-model-based steganography, achieving provable security (zero KL divergence) while dramatically improving embedding capacity over previous provably secure methods. Prior approaches sacrificed capacity for security guarantees. Combined with the week's other agent steganography papers (Undetectable Conversations, ACF), this represents a convergent research push on covert AI communication channels.