S-Stega: Restoring Temporal Stability in Generative Language Steganography via Salt Randomization

ACL ARR 2026 January Submission6224 Authors

Keywords: Linguistic steganography, Prefix leakage, Pseudorandom salting

Abstract: We propose \textbf{S-Stega}, a salt-randomized preprocessing framework for generative steganography that mitigates temporal instability arising from fixed or semi-fixed bit-consumption schemes. We show that deterministic encoders can induce reproducible, prefix-aligned regularities—manifesting as prefix leakage and time-structured drift—thereby motivating the \textbf{Chosen-Secret-Message Adversary (CSMA)} and an \textbf{IND-CSMA} security notion. Under ideal random and pseudorandom assumptions, we prove that salting masks message bits in both statistical and computational senses, suppressing prefix-aligned leakage and attenuating temporal drift. In practice, \textbf{S-Stega} XORs a domain-separated pseudorandom salt into the bit stream prior to embedding, preserving lossless decoding and providing a codec-agnostic hardening layer for HC, AC, and ANS. Experiments across multiple LMs and four datasets show that salting drives prefix agreement near zero, weakens short-range temporal dependence, reduces trajectory separability by 40\%--70\%, and lowers time-sensitive steganalysis accuracy by roughly 30\%, while largely preserving text quality with method-dependent effects on embedding rate.

Paper Type: Long

Research Area: Language Models

Research Area Keywords: security and privacy

Contribution Types: Model analysis & interpretability, Approaches low compute settings-efficiency, Data analysis

Languages Studied: English

Submission Number: 6224