AB
AB
AiBoss

All Categories

project

SynthID Text - Google DeepMind's AI-generated text watermarking technology

SynthID Text is a text watermarking technology developed by Google DeepMind, used to identify and verify text generated by large language models (LLMs). It relies on subtle adjustments to the token probability score embedding during the generation process, making it virtually imperceptible...

What is SynthID Text?

SynthID Text is a text watermarking technology developed by Google DeepMind for recognizing and verifying text generated by large language models (LLMs). It embeds a virtually imperceptible watermark by subtly adjusting the token probability scores during the generation process, achieving high detection accuracy without compromising text quality or user experience. SynthID Text is based on the Tournament sampling algorithm, supports both lossless and lossy modes, and can be implemented in large-scale production systems with extremely low additional computational overhead. The technology has been successfully applied to the Gemini and Gemini Advanced systems, providing new possibilities for using AI technologies.

SynthID Text's main functions

  • Text watermark embeddingIt can embed digital watermarks into text generated by large language models (LLMs), and the watermarks are used to identify and verify the source of the text.
  • Quality maintenanceThe goal is to maintain the original quality and natural flow of the text during the watermark embedding process, without affecting the reading experience.
  • High detection accuracyThe watermark design allows embedded watermarks to be detected efficiently, identifying whether the text was generated by a specific LLM.
  • Minimize latencyThe watermarking process is designed to have minimal impact on the latency of the generated text, making it suitable for real-time or large-scale text generation scenarios.
  • It does not affect LLM training.The watermarking process is modified only during the sampling stage of text generation and does not affect the model training process.

The technical principles of SynthID Text

  • Sampling algorithm modificationSynthID Text embeds watermarks using a modified sampling algorithm of a Large Language Model (LLM). While generating text, the model selects the next token based on a probability distribution; SynthID Text adjusts the probabilities during this process to embed a barely perceptible watermark.
  • Tournament samplingSynthID Text uses the Tournament sampling algorithm, which selects a token using a simulated tournament process. In each round, multiple tokens are randomly selected, and a winner is chosen based on a score associated with a watermark function. This process is repeated multiple times until a final token is selected as the output.
  • Random seed generationAt each step of text generation, SynthID Text requires a random seed to influence the choice of token. The seed is generated based on the previous text and watermark key, ensuring the randomness and unpredictability of the watermark.
  • Non-distortion configurationSynthID Text can be configured in a lossless mode, embedding watermarks while maintaining text quality. The watermark embedding does not affect the original probability distribution of the text, ensuring the naturalness and coherence of the text.
  • Watermark detectionDuring the detection phase, SynthID Text determines whether the text contains a watermark based on calculated statistical features (i.e., the g-value). If the text contains a watermark, the statistical features will show a deviation consistent with the watermark function.

SynthID Text project address

Application scenarios of SynthID Text

  • Content verificationIn the fields of news, publishing, and academia, this involves verifying whether articles are written by humans or generated by AI to ensure the credibility of the content.
  • Educational assessmentIn educational settings, this technology helps detect whether student assignments are generated using AI-generated text, thus upholding academic integrity.
  • Law and complianceIn the legal field, S detects whether legal documents or contracts contain AI-generated text, ensuring the legality and validity of the legal documents.
  • social mediaOn social media platforms, it helps identify and tag AI-generated content to prevent the spread of misleading information.
  • Customer ServiceIn automated customer support systems, the distinction between automatically generated responses and human responses is made to improve customer service quality.
Back