Language steganography drops the bit-exact requirement: STEALTH decodes every message across precisions, GPUs and even model families
Synopsis
Provably secure linguistic steganography has assumed the receiver can reproduce the sender's token probabilities bit for bit, yet floating-point precision, quantization, GPU architecture, or scoring a text in one pass instead of token by token all break that premise. STEALTH samples each token with a keyed exponential race, so a probability mismatch flips only the token whose race is close, and a rateless spinal code decoded by beam search absorbs those flips. Across the Gemma, Llama and Qwen families, same-family receivers decoded every message at every precision, quantization level, smaller model size and on a second GPU, and cross-family receivers decoded every message tested.
Interpretation
Linguistic steganography no longer needs sender and receiver to compute identical token distributions; a receiver can recover the hidden message while differing in precision, quantization, model size or GPU architecture. Earlier provably secure schemes encoded the message in exactly where each sampled token fell within the distribution, so the receiver had to reproduce the sender's distribution bit for bit to invert it. With Gemma-2-9B-it, Llama-3.1-8B-Instruct and Qwen2.5-7B-Instruct as senders, arithmetic coding and Discop decoded every message when replaying the sender's own path; both decoded none when replaying on a different GPU, quantizing to int8 or NF4, or substituting a smaller same-family model, and arithmetic coding also decoded none when scoring the same model in one pass, while Discop decoded only 0.5–1.5% that way.
The sender draws noise for each token from a keyed exponential race, so a receiver with a slightly different distribution picks the same token almost everywhere and a mismatch costs only that position's evidence. The exponential race compares each token's score against independent noise instead of carrying state across positions as interval coding does, so one wrong pick does not corrupt every later token. Messages are 32-bit payloads with a 16-bit CRC, spine values are assigned by counting words rather than tokens, and the receiver teacher-forces the whole text once before beam search; over 200 CNN/DailyMail articles and 200 messages per condition, every same-family configuration reached 200 of 200, with the worst case, a Gemma 2B receiver, needing no more than 384 tokens.
Perspective
The demonstrated setting is 7–9B senders from three model families, 32-bit payloads over 200 CNN/DailyMail articles, and receivers spanning same-family bf16, fp16, fp32, int8 and NF4, smaller same-family models, and the other two model families, on NVIDIA A6000 and H100 hardware.
The security proof assumes HMAC-SHA256 and ChaCha20 behave as pseudorandom functions; the sender's own decodability check is reliable only for receivers close to its configuration, and cross-family receivers decode 12.5–90.5% of messages at the sender's stopping point, needing more text.
