SynthID is a technology developed by Google DeepMind for applying Invisible Watermarking to AI-generated content, primarily text and images. It aims to help identify AI-generated media reliably without affecting the user experience.
Key Features
- Text Watermarking: Embeds a hidden digital signature into generated text by subtly adjusting token probabilities during generation, based on a secret key. This process doesn’t require model retraining and is integrated into libraries like Hugging Face Transformers via logits processors.
- Image Watermarking: Applies an imperceptible watermark directly into the pixels of AI-generated images.
- Detection: Specialized detectors, using the corresponding secret key for text, can identify the presence of the SynthID watermark with a certain probability (detection is often probabilistic, yielding “watermarked,” “not watermarked,” or “uncertain”).
- Open Source & Integration: The text watermarking component has been open-sourced and integrated into popular AI development tools. Image watermarking/detection is available via Google Cloud.
Purpose
- AI Content Identification: Provides a technical means to distinguish AI-generated text and images from human-created ones.
- Responsible AI: Part of Google’s toolkit for promoting responsible AI practices by enabling AI Content Labeling and Content Provenance.
Connections
- Is a type of Invisible Watermarking
- Developed by Google DeepMind (part of Google)
- Used for AI Content Labeling and Content Provenance
- Supports AI Transparency Requirements
- Integrated with Hugging Face Transformers for text generation