RRC Steganography

Rotation Range-Coding (RRC) Steganography — an efficient and provably secure linguistic steganographic method that embeds secret messages into natural-language text generated by large language models.

Paper: Efficient Provably Secure Linguistic Steganography via Range Coding

How It Works

The rotation mechanism solves two issues of vanilla range-coding steganography:

1. Probability-distribution preservation — the rotated secret value is uniformly distributed within each interval, so the token selection follows the original LM distribution.
2. Fresh randomness — a new PRNG offset at each step prevents randomness reuse.

Repository Structure

.
├── RRC_embed.py   # Embedding (Algorithm 3) 
├── RRC_extract.py   # Extraction (Algorithm 4) 
├── test_roundtrip.py         # Self-contained end-to-end verification script
├── requirements.txt          # Python dependencies
└── README.md

Requirements

• Python ≥ 3.10
• PyTorch ≥ 2.0
• Transformers ≥ 4.35

Install dependencies:

pip install -r requirements.txt

Quick Start

1. Verify correctness

The round-trip test uses meta-llama/Llama-2-7b-hf and runs on CPU/CUDA:

python test_roundtrip.py

Expected output:

>>> Step 3: Verification
    Original: 10110011001010111010011100101011...
    Extracted: 10110011001010111010011100101011...

    ✅ SUCCESS — extracted message matches perfectly!

2. Embed a secret message

Prepare a prompts file 0.Prompts.tsv with columns idx and text, then run:

python RRC_embed.py \
    --language_model meta-llama/Llama-3.1-8B \
    --bit_length 128 \
    --key 42 \
    --top_k -1

3. Extract the secret message

python RRC_extract.py \
    --language_model meta-llama/Llama-3.1-8B \
    --bit_length 128 \
    --key 42 \
    --input_file 1.RC_decimal_Llama-3.1-8B_bit128.tsv

⚠️ Important: The sender and receiver must use the same language model, key, bit length, and top-k setting for correct extraction.

CLI Arguments

RRC_embed.py

RRC_extract.py

Technical Note: Decimal Modulo

Python's Decimal type uses truncated division for the % operator, which can return negative remainders (e.g. Decimal('-19672') % Decimal('65536') → -19672 instead of 45864). This implementation uses a decimal_mod helper that guarantees the result is always in [0, m):

def decimal_mod(a, m):
    return a - m * (a / m).to_integral_value(rounding=ROUND_FLOOR)

⚠️ For reliable secret extraction, some operations (sort, cumsum) are offloaded to CPU at float64 precision to avoid CUDA non-determinism. This trades a amount of speed for guaranteed encode–decode consistency.