SynthID – Watermarking and identifying AI-generated text
summary
SynthID-Text implements a lightweight watermark that embeds a secret key into generated text by computing per‑n‑gram bits and using two‑token tournament sampling with multiple tournament layers to bias choices toward watermark value 1. It is open source in HuggingFace Transformers and ships with a procedure to configure keys, call model.generate, and train a Bayesian detector on watermarked vs. clean corpora. Empirical results show no significant drop on benchmarks and no preference loss in human studies, and detection looks for significantly more 1's than 0's.
AI-written summary · about
related talks
44mDigital watermarking for generative AI
27mTruth or Fiction The Art of Unveiling DeepFakes in Text, Audio, Image
45mTips For Detecting The Use Of AI-Generated Text
27mTurning Your Weakness Into a Strength: Watermarking Deep Neural Networks by Backdooring
55mAI Showdown: The Good, The Glitchy, and The Risky
56mHomoglyph-Based Attacks: Circumventing LLM Detectors
40mAI vs AI: Creating Novel Spam and Catching it with Text Generating AI
46mTackling Deepfakes, Wars, and Other Security Threats in the GenAI World