Auto-Stega: An Agent-Driven System for Lifelong Strategy Evolution in LLM-Based Text Steganography

Jiuan Zhou, Yu Cheng, Yuan Xie, Zhaoxia YinarXiv 2025 · DOI · arXiv

Introduction

Steganography is the practice of concealing secret information within a non-secret "cover" medium. In the context of Large Language Models (LLMs), text steganography aims to hide data within generated text. However, researchers face a persistent "trilemma" in balancing three competing objectives: embedding rate (data capacity), imperceptibility (naturalness to humans), and security (resistance to AI-powered steganalysis).

Traditional methods often rely on static, hand-crafted rules that struggle to maintain this balance, particularly at high embedding rates. These methods frequently produce text that appears unnatural or exhibits statistical anomalies detectable by advanced classifiers. Furthermore, static algorithms lack the adaptability required to handle diverse linguistic domains, such as transitioning from formal news reports to informal social media posts.

System Architecture

To address these challenges, we introduce Auto-Stega, the first agent-driven, self-evolving framework for text steganography. Auto-Stega treats steganography as a lifelong learning task, utilizing a closed-loop system of LLM agents to autonomously discover, evaluate, and refine hiding strategies.

The Auto-Stega framework architecture, featuring an agent-driven loop for strategy evolution and the PC-DNTE encoding algorithm.

The framework consists of several specialized components:

  • Web Searcher Agent: Continuously mines research literature to identify candidate steganographic methods, converting them into executable strategy entries.
  • Strategy Library: A dynamic database that stores successful strategies, categorized by their applicable scenarios (e.g., news, movies, or tweets).
  • Steganography LLM: Acts as the "sender," retrieving optimal strategies from the library to generate stego text that embeds the secret information.
  • Scorer & Summarizer Agents: These agents evaluate the generated output based on quality and security metrics. The Summarizer distills these evaluations into "lessons learned," which are used to update and evolve the Strategy Library over time.

The PC-DNTE Algorithm

A critical technical innovation within Auto-Stega is the PC-DNTE (Parity-Constrained Dynamic Nucleus-Typical Encoder). This plug-and-play algorithm is designed to achieve high embedding rates while maintaining strict alignment with the base LLM’s natural probability distribution. By using dynamic "bins" and parity checks during token selection, PC-DNTE ensures that the generated text remains statistically indistinguishable from normal LLM output, thereby maximizing security against eavesdroppers.

Lifelong Strategy Evolution

The core strength of Auto-Stega lies in its ability to move beyond static algorithms toward a model of continuous improvement. By allowing agents to analyze why specific strategies fail and iteratively update the library, the system achieves a form of "lifelong evolution."

Illustration of the lifelong evolving steganographic strategy, showing how strategies are refined and scored over time.

This evolutionary process allows Auto-Stega to compose multiple strategies to meet specific constraints, such as prioritizing high security in low-capacity environments. The system's success demonstrates that preserving the entire statistical distribution of the generated text is the "holy grail" of high-capacity steganography.

Experimental Evaluation

Auto-Stega was evaluated across three diverse public corpora to test its adaptability and stylistic consistency: the News Category Dataset (formal headlines), the Large Movie Review Dataset (opinionated prose), and Sentiment140 (informal tweets).

Key Results

The framework significantly outperformed state-of-the-art (SOTA) methods like ADG and Discop:

  • Naturalness: At high embedding rates, Auto-Stega achieved a 42.2% reduction in perplexity deviation, indicating text that is much closer to natural cover text.
  • Security: The system showed a 1.6% improvement in anti-steganalysis performance against advanced detectors like LS-CNN and BERT-FT.
  • Human Preference: Blind evaluations confirmed that humans consistently find Auto-Stega's output more fluent, coherent, and contextually relevant than baseline methods.

Conclusion and Future Work

Auto-Stega represents a paradigm shift in text steganography by replacing static rules with an evolving, agent-driven architecture. While the current system is highly effective for text, future research will explore extending the framework to multimodal covers (e.g., images and text) and optimizing the agent-driven loop to reduce inference latency for real-time applications. Additionally, further testing is required to ensure robustness against noisy communication channels where text may undergo minor formatting changes during transmission.

Summary updated

BibTeX
@misc{zhou2025auto,
  author = {Jiuan Zhou and Yu Cheng and Yuan Xie and Zhaoxia Yin},
  title = {{Auto-Stega: An Agent-Driven System for Lifelong Strategy Evolution in LLM-Based Text Steganography}},
  year = {2025},
  eprint = {2510.06565},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.06565},
}
Figure