Google's SynthID for Multi-Modal Content

Google DeepMind's SynthID embeds imperceptible digital watermarks directly into AI-generated images, audio, text, and video, designed to withstand modifications like cropping or compression and detectable via a public portal or APIs.

The top 3

  1. How SynthID's Invisible Mark Differs from Old Watermarks: Unlike traditional watermarking that adds visible layers or easily removed metadata, Google's SynthID embeds an imperceptible digital signature directly into content pixels, audio waves, or text token probabilities during AI generation, using co-trained neural networks for robust detection.
  2. SynthID's Multi-Modal Reach: Images, Audio, and Text: Initially for images generated by Google's Imagen, SynthID has expanded to watermark audio from models like Lyria and NotebookLM, and text from Gemini. While effective across modalities, its robustness against manipulation, such as paraphrasing or translation, is significantly stronger for images and audio than for text.
  3. How SynthID Defies Common Content Manipulation: SynthID's robustness comes from distributing the watermark across the entire content and using sub-perceptual changes, allowing it to survive common edits like JPEG compression (down to 70%), moderate cropping (up to 30-40%), format conversions, and brightness adjustments for images.

Sources

Open the full topic