HiTMS Batches Eight Steganographic Streams Through One LLM Call for 4.3x Throughput and Lower Detectability

arXiv 2607.23597·low signalXBlueskyLinkedInCopy link

Generative linguistic steganography hides secret bits in an LLM's sampling randomness, but every existing scheme is single-stream — one secret, one prompt, one response — which offers no protocol-level support for batched inference and leaks slot occupancy when naively co-batched. HiTMS (arXiv 2607.23597) distributes a secret across multiple responses produced jointly over successive rounds, embedding and extracting several streams per batched call, wrapping each response in a self-describing frame and using a key-derived schedule that binds streams to slots and fills unused slots with decoys. Across eight dataset-model-coder settings, eight-stream HiTMS reaches up to 4.3x faster embedding and extraction than single-stream baselines while dropping average steganalyzer AUROC from 0.681 to 0.601, with sustained gains from 4 to 64 streams. Code at github.com/ryehr/HiTMS_steganography.