AI Watermark Robustness Lab

Watermark Generator & Token Config

Prompt / Generated Text

Greenlist Ratio (γ): 0.50

Logit Bias (δ): 2.0

Hash Seed Key

Detection Threshold (α)

Evasion & Perturbation Attack

Attack Vector

Perturbation Rate: 35%

Statistical Detection Metrics Live Testbench

Clean Z-Score

5.84

Green: 76.7%

Attacked Z-Score

0.73

Green: 53.3%

Entropy Delta

+3.8%

T = 120 tokens

Attacked Watermark Status:

Watermark Evaded (Inconclusive, z < 2.33)

Gaussian Hypothesis Distribution (Null H₀ vs Watermark H₁)

Under unwatermarked null hypothesis H₀, expected greenlist count is 60.0 (50%). Clean stream contains 92 greenlist tokens (76.7%, z = 5.84, p < 0.0001). Under 35% synonym attack, greenlist count degrades to 64 (53.3%, z = 0.73), evading EU AI Act threshold.

Interactive Token Stream & Cryptographic Hash Provenance

Greenlist Token

Redlist Token

Attacked/Perturbed

Hover or click any token in the ledger to inspect its previous-token hash, pseudo-random greenlist partition, and perturbation status.

LLM Watermarking & Evasion Mechanics

Read the explanation

Kirchenbauer watermarking uses previous token hashes to split candidate vocabularies into green and red partitions, applying a logit bias delta so green tokens dominate generation. During verification, a statistical z-score evaluates green count versus null expectation. Clean watermarked text easily clears the detection threshold alpha with high statistical confidence. Increasing the attack perturbation rate replaces greenlist tokens with synonyms or homoglyphs. This degrades the green ratio back toward fifty percent, causing z to collapse below threshold into evasion.