🔐 AI Text Watermarking

Invisible, Contextual, and Neural Watermarking for AI-Generated Text

This repository contains a complete implementation of invisible watermarking techniques for AI-generated text, including both Unicode-based watermarking and a more advanced contextual neural watermarking system using PyTorch and HuggingFace Transformers.

The goal of this project is to embed a watermark during generation so that text remains untouched visually but can be reliably detected later — even after copy/paste or light editing.

---

📌 Features

✅ 1. Invisible Unicode Watermarking (Basic Method)

• Uses zero-width Unicode characters
• Fully invisible to users
• Survives copy/paste into plain editors
• Good baseline watermarking method

✅ 2. Contextual Neural Watermarking (Advanced Method)

Built using state-of-the-art techniques:

• EnhancedHashNet: Neural context-based hashing
• Green/Red token lists per decoding step
• Logit manipulation using custom LogitsProcessor
• Dynamic watermark embedding
• Statistical detection using Z-score and p-values
• Visualization of watermark patterns

This method provides higher security, robustness, and stealth.

---

🧠 How It Works

🔹 Watermark Encoding

During text generation:

1. Context Analysis: The model takes previous tokens (context window)
2. Neural Hashing: A neural hash network generates a unique seed based on context
3. Vocabulary Permutation: Vocabulary is permuted using this seed
4. Token Biasing: "Green tokens" are boosted, "red tokens" are penalized
5. Natural Selection: Model is more likely to choose green tokens → creates invisible pattern

The watermark is embedded seamlessly without affecting text quality or fluency.

🔹 Watermark Detection

Given a text to verify:

1. Seed Reconstruction: Detector reconstructs the seed at each position
2. Green-list Rebuilding: Rebuilds green-lists exactly as during generation
3. Pattern Matching: Checks how often text chooses green tokens
4. Statistical Analysis: Computes:
   • z-score: Measures deviation from random selection
   • p-value: Statistical significance of watermark presence
   • confidence: Overall detection confidence score

---

🚀 Installation

Prerequisites

• Python 3.8+
• PyTorch 2.0+
• HuggingFace Transformers
• NumPy, SciPy

Install Dependencies

pip install torch transformers numpy scipy matplotlib

Or use requirements.txt:

pip install -r requirements.txt

---

🔧 Configuration

Watermark Parameters

Parameter	Description	Default
context_width	Number of previous tokens used for hashing	5
gamma	Proportion of vocabulary marked as "green"	0.25
delta	Logit bias added to green tokens	2.0
detection_threshold	Z-score threshold for detection	4.0

Tuning Guidelines

• Higher gamma: More tokens marked green → stronger watermark, potentially less natural
• Higher delta: Stronger bias → more detectable but may affect quality
• Larger context_width: More secure but slower detection

---

📊 Detection Metrics

The detector provides several metrics:

• Z-score: Measures how unusual the green token frequency is

  • z > 4.0: Strong watermark detected
  • 2.0 < z < 4.0: Weak signal
  • z < 2.0: No watermark

• P-value: Probability of observing this pattern by chance

  • p < 0.0001: Very high confidence
  • p < 0.05: Significant detection

• Green Token Ratio: Percentage of tokens that are green

  • Expected ratio without watermark: gamma (e.g., 0.25)
  • With watermark: typically > 0.5

---

🛡️ Robustness

What the Watermark Survives

✅ Copy/paste operations
✅ Light paraphrasing
✅ Minor edits
✅ Format changes

Limitations

❌ Heavy rewriting or summarization
❌ Translation to another language
❌ Adversarial attacks specifically designed to remove watermarks

---

📈 Visualization

Generate watermark pattern visualizations to analyze the detection results and see the distribution of green tokens throughout the text.

---

📚 References

This implementation is based on research in:

• "A Watermark for Large Language Models" (Kirchenbauer et al., 2023)
• "On the Reliability of Watermarks for Large Language Models" (Christ et al., 2023)
• Zero-width character steganography techniques

---

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.