Token-Probability Watermarks
SynthID embeds signals during generation by hashing n-gram context ($k=4$) to bias token selection toward a green-list vocabulary — not metadata, not visible text.
Gemma 4 E2B Engine Live
Break AI Text Watermarks.
An adversarial ML pipeline with confidence-based attack selection — Gemma 4 E2B, perplexity detection, homoglyphs, and token perturbation.
Try the Live Tool NowWhat This Demonstrates
ML KnowledgeNot just an API wrapper — this project shows understanding of LLM watermarking at the token-probability level.
G-Value as Statistical Measure
Mean green-list alignment above 0.55 indicates watermark confidence; ~0.49–0.51 is baseline noise. The workspace shows pre/post G-values on every run.
Why Paraphrasing Works
Regenerating text with an unwatermarked model (Gemma 4 E2B) draws fresh token sequences — n-gram hash patterns reset regardless of original watermark strength.
Character vs Statistical Attacks
Unicode tricks (zero-width chars, homoglyphs) break character-level watermarks; paraphrase and perturbation target statistical logit biasing — different physics, different defenses.
Self-Evaluating Pipeline
Every attack is scored before and after: G-value drop %, perplexity, auto-selector rationale, and a 5-step transformation timeline prove effectiveness quantitatively.
See Live Scores →Attack Strategies
Mode 1: Combined Paraphrase + Perturb
1. Sanitize →
2. Pre-Detect →
3. Paraphrase →
4. Perturb
IN
Google DeepMind SynthID technology embeds an imperceptible statistical watermark into AI-generated text by biasing token selection probabilities during logit processing.
OUT
Success
Signal Drop: ~31%
DeepMind's SynthID framework incorporates a subtle statistical signal into AI text by altering token sampling distributions. Paraphrasing cleanly erases this signature.
MATHEMATICAL FORMULATION
Method Artifacts
Animated pipeline previews for core detection and attack methods.
Provider Benchmarks
Tech: SynthID N-gram Logit Biasing
Google Gemini uses SynthID n-gram context logit biasing. Watermarks are embedded into token selection probabilities during model sampling and survive formatting or minor edits.