PythonMachine Learning
Adversarial ML
Gemma 4 E2BG-Value Scoring
Watermark Detection
SynthID Watermark Remover is an adversarial ML pipeline designed to detect and erase SynthID statistical watermarks from LLM-generated text. Rather than acting as a simple API wrapper, this project demonstrates token-probability level watermark physics. It measures pre- and post-mitigation G-values, routing text through a confidence-based selector that applies targeted adversarial attacks—including perplexity detection, homoglyphs, and token-probability perturbations—to completely erase the statistical watermark signature while maintaining natural readability.
Features
- •Token-Level Detection: Employs G-value scoring to measure statistical green-list alignment confidence vs. baseline noise (0.49-0.51).
- •Confidence-Based Selector: Dynamically routes adversarial attacks based on detection confidence and perplexity thresholds.
- •Watermark Physics Erasure: Implements token perturbation and homoglyph substitutions to disrupt the hidden hashing parameters.
- •Gemma 4 E2B Engine Live: Demonstrates real-time watermark insertion and removal using Google's Gemma models.
- •Before/After Metrics: Displays quantitative comparisons proving successful watermark removal with minimal text quality degradation.
- •Retro Brutalist UI: Features a custom responsive workspace styled with neon banners, bold outlines, and interactive dashboards.