TI-StegoAlign: Channel-Guided Post-Training for Generative Text Steganography under Tokenization Inconsistency

2026-07-31

AI Summary

This work addresses the critical challenge in generative text steganography where decoding failures frequently occur due to tokenization mismatches between sender and receiver, leading to desynchronization of the secret message extraction process. To resolve this issue, the authors propose a channel-guided post-training framework that, for the first time, explicitly incorporates the receiver’s tokenization channel characteristics into steganographic alignment training. The approach eliminates the need for on-the-fly tokenization correction during communication and achieves high robustness through lightweight LoRA fine-tuning. By integrating a Bit Consistency Supervision Objective (BCSO) with Channel-Conditioned Preference Optimization (CCPO), the method simultaneously preserves high textual quality and enhances resistance against steganalysis. Experimental results demonstrate that the proposed method achieves 100% bit accuracy at the receiver end, reduces normalized perplexity deviation by 21.6% compared to the strongest baseline, and improves steganalysis resistance by 6.3%.

Abstract

Generative text steganography enables LLM agents to exchange secret information through task-relevant messages. Yet most methods evaluate recovery on sender-side tokens, whereas the receiver observes only surface text. Detokenization and receiver-side retokenization can alter token boundaries, desynchronize coding states, and cause such evaluation to overestimate receiver-side recovery. Existing remedies rely on inference-time filtering or verification, correcting individual outputs without adapting the generation policy to the receiver-side channel. To address these limitations, we propose TI-StegoAlign, a channel-guided post-training framework. The Bit-Consistent Supervised Objective (BCSO) enlarges local coding margins at realized sender-side embedding positions. Channel-Conditioned Preference Optimization (CCPO) then aligns complete stegotexts using receiver-realistic recovery, text quality, and anti-steganalysis feedback. TI-StegoAlign updates only LoRA parameters and requires no tokenization-specific correction during communication. Experimental results show 100% receiver bit accuracy. Compared with the strongest baselines, TI-StegoAlign achieves a 21.6% reduction in normalized perplexity deviation and a 6.3% relative improvement in anti-steganalysis performance.

Problem

• text steganography
• tokenization inconsistency
• generative models
• receiver-side recovery
• LLM agents

Innovation

• text steganography
• tokenization inconsistency
• channel-guided post-training
• LoRA adaptation
• preference optimization