Stego Arithmetic Coding

A Python library for steganographic encoding and decoding using arithmetic coding to hide secret messages within LLM-generated text.

Features

• Arithmetic Coding Steganography: Hide messages by influencing token selection in language models
• OpenAI Integration: Uses OpenAI's API to get token probabilities
• CLI Interface: Command-line tools for encoding, decoding, helper conversions, and visualization
• Artifacts & Visualization: Optional artifacts (JSON logs) and SVG visualizations for each step
• Flexible API: Easy-to-use Python API for integration

Installation

# Basic installation
pip install .

# With CLI support
pip install .[cli]

# With testing dependencies
pip install .[test]

Quick Start

Python API

from stego_arith_coding import ArithmeticSteganography

# Initialize with your OpenAI API key
stego = ArithmeticSteganography(model="gpt-3.5-turbo-1106") # WARN: Other models are non-deterministic

# Encode a message
context = "The weather today is "
message = "SECRET"
message_bits = stego.text_to_bits(message)
encoded_text = stego.encode(context, message_bits)

print(f"Encoded: {encoded_text}")

# Decode the message
decoded_bits = stego.decode(context, encoded_text, message_length=len(message_bits))
decoded_message = stego.bits_to_text(decoded_bits)

print(f"Decoded: {decoded_message}")

CLI Usage

Global options:

• --api-key Override OPENAI_API_KEY for this invocation
• -v, --verbose Enable verbose logging

# Set your OpenAI API key (or pass --api-key)
export OPENAI_API_KEY="your-api-key-here"

# Encode a message (optionally write artifacts to a directory)
stego-arith encode "The weather today is " "SECRET" -a artifacts/encode_run

# Decode a message (you can pass the text directly or via a file)
stego-arith decode "The weather today is " "sunny and warm"
stego-arith decode "The weather today is " --input artifacts/encode_run/encoding-result.txt -a artifacts/decode_run

# Convert text to binary (with ETX EOF marker)
stego-arith text-to-bits "Hello"

# Convert binary to text (stops at ETX EOF marker)
stego-arith bits-to-text "0100100001100101011011000110110001101111"

# Render an SVG visualization from a coding_data.json file
stego-arith visualize artifacts/encode_run/encoding-coding-data.json

Artifacts:

• When -a/--artifacts DIR is provided, the CLI writes step-by-step JSON and logs to DIR/:
  • encoding-coding-data.json or decoding-coding-data.json
  • encoding-visualization.svg or decoding-visualization.svg (auto-rendered if JSON is present)
  • encoding-verbose.log or decoding-verbose.log
  • encoding-result.txt or decoding-result.txt

How It Works

The library uses arithmetic coding steganography to hide messages:

1. Context Analysis: Gets token probabilities from a language model for the current context
2. Interval Division: Divides the probability space into intervals based on token probabilities
3. Message Encoding: Selects tokens based on the secret message bits to encode
4. Synchronization: Maintains perfect synchronization between encoding and decoding

The key insight is that by carefully selecting which tokens to generate based on their probabilities, we can encode arbitrary binary data while maintaining natural-looking text.

Visualization

Example of the encoding process is below for: stego-arith encode "The weather today is " "S"

API Reference

ArithmeticSteganography

Main class for steganographic operations.

Methods

• encode(context, message_bits, precision=16, artifacts_dir=None): Encode message bits into text; optionally write artifacts to artifacts_dir
• decode(context, encoded_text, precision=16, message_length=None, artifacts_dir=None): Decode message from text; optionally write artifacts to artifacts_dir. Decoding uses an ETX (0x03) EOF marker; message_length is optional.
• text_to_bits(text): Convert text to binary representation with ETX EOF marker appended
• bits_to_text(bits): Convert binary data back to text; stops at the first ETX EOF marker

Requirements

• Python >= 3.11
• OpenAI API key
• Dependencies: openai, tiktoken

Testing

# Run tests (requires OPENAI_API_KEY for integration tests)
python -m pytest tests/

# Run specific test file
python tests/test_core.py

FAQ / Issues

See limitations in my blog post at https://www.artkpv.net/Tool-Arithmetic-Coding-for-LLM-Steganography/

References

• Ziegler, Zachary, Yuntian Deng, and Alexander Rush. ‘Neural Linguistic Steganography’. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), edited by Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan. Association for Computational Linguistics, 2019. https://doi.org/10.18653/v1/D19-1115.
• Pasco, ‘Source Coding Algorithms for Fast Data Compression’.

License

MIT License - see LICENSE file for details.