LLM Statistical Text Watermarking & Robustness Lab

1. Watermark Hyperparameters & Input

Partition Ratio (γ): 0.50

Logit Boost (δ): 2.0

Hash Seed Key:

Context N-Gram:

Analyzed Text Passage (Interactive Edit):

Token Classification Stream (0 tokens)

Click token to inspect cryptographic partition

Selected Token Inspector: Click any token chip above to view hashing state.

2. Statistical Provenance Detection Z-Score Detector

Total Tokens (N)

Green Tokens (|G|)

Expected Null (γ·N)

18.5

Green Ratio

83.8% green ratio

Detection Classification:

Watermark Confirmed (p < 0.0001)

One-Tailed Z-Score:

4.11

Null Hypothesis Distribution vs Observed Signal

← Natural Text Domain (Z ≤ 2.0) Watermarked Domain (Z > 3.0) →

Adversarial Robustness Summary

Under Kirchenbauer et al., each previous token acts as a cryptographic seed to pseudo-randomly split the vocabulary into green and red lists. At generation time, green tokens receive a logit bias boost (+δ). As text length N increases, detection confidence scales rapidly: \(Z = \frac{|G| - \gamma N}{\sqrt{N\gamma(1-\gamma)}}\).