SynthID - DeepMind launches tools that can embed digital watermarks and detect AI-generated content.

What is SynthID?

SynthID is a technology tool developed by DeepMind that helps identify content by embedding digital watermarks in AI-generated content. Watermarks are imperceptible to humans but detectable by machines, enhancing the trustworthiness of information. SynthID is suitable for various content formats, including text, music, images, and videos, and maintains detectability even after content has been modified or compressed. Currently in testing, SynthID is gradually being integrated into various products, supporting AI-generated content from individuals and organizations. By adjusting probability scores in generated text or adding watermarks to audio waveforms, image pixels, and video frames, SynthID ensures the originality and authenticity of content without compromising quality or creativity.

SynthID's main functions

• Watermark EmbeddingThis involves embedding digital watermarks directly into AI-generated content, such as text, music, images, and videos, without compromising the quality of the original content.
• Content recognitionIt scans various media formats to detect the presence of digital watermarks, helping users confirm whether content was generated by AI tools.
• Cross-media applicationsSuitable for different media types, including text, music, still images and video, providing a wide range of application scenarios.
• Resistance to modificationThe watermark is designed to be detectable even after the content has been modified, such as by cropping, adding filters, adjusting colors, changing frame rates, or using lossy compression.
• Enhance trustBased on the recognition of AI-generated content, it helps improve the credibility of information and combat misinformation and inappropriate content attribution.

SynthID's technical principles

• Deep learning models: Using deep learning models and algorithms to embed and recognize watermarks.
• Probability Adjustment(For text): During the text generation process, adjusting the probability score of the next token predicted by the language model to embed the watermark will not affect the coherence and quality of the text.
• Spectrum processing(For music): Convert the audio waveform to a spectrogram, add a watermark, and then convert it back to a waveform to ensure that the watermark is not visible to human hearing.
• Pixel-level embedding(For images and videos): Watermarks are embedded directly into the pixels of an image or the frames of a video, and can remain detectable even after various modifications.
• Robust designThe watermark design takes into account various possible content modifications to ensure reliable detection under different conditions.

SynthID project address

• Project official website:deepmind.google/technologies/synthid

Application scenarios of SynthID

• Content verificationIn the news and media industry, this involves verifying whether news photos, videos, or audios were generated by AI to ensure the authenticity of reports.
• Copyright protectionFor artists and content creators, embedding watermarks protects the copyright of their works and prevents unauthorized use and distribution.
• Education and ResearchIn academic research, it helps identify whether research materials are generated by AI, ensuring the reliability and effectiveness of the research.
• Law and complianceThe legal industry uses AI to detect whether evidence has been tampered with, thus upholding judicial fairness.
• social mediaOn social media platforms, it helps users identify and filter out fake news or misleading content generated by AI.