Discop's core algorithm is modality-agnostic and deploys unchanged across text generation (GPT-2, DistilGPT-2, Transformer-XL), image completion (Image GPT), and text-to-speech (Tacotron + WaveRNN), requiring only that both parties share the generative model, PRNG, and seed. The same zero-KLD security proof applies across all modalities.
From 2023-ding-discop — Discop: Provably secure steganography in practice based on ``distribution copies''
· §IV
· 2023
· Symposium on Security \& Privacy
Implications
Implement generative steganography as a modality-agnostic layer above any explicit autoregressive model so the circumvention tool can rotate cover media (text posts, synthetic images, TTS audio) in response to selective blocking without redesigning the covert-channel protocol.
Distributing cover-medium selection across multiple generative modalities (text, image, speech) forces the censor to block all autoregressive AI output at once, imposing prohibitive collateral damage on legitimate traffic.