DeepMind's SynthID watermarks LLM-generated text
A statistical signature that doesn't change the meaning of LLM-generated text.
DeepMind has released SynthID Text, a tool that can help detect text that was created with LLMs. SynthID uses watermarking techniques to make modifications to the text in a way that keeps its meaning while making it detectable by a classifier model.
SynthID uses “generative modeling,” a class of watermarking techniques that modify the sampling stage of the LLM. This means that you don’t need to retrain the model, which makes the technique resource-efficient. SynthID makes context-specific changes to the generated text to create a statistical signature in the generated text while maintaining its quality. SynthID has a set of parameters that can be configured separately for different LLMs.
Each SynthID configuration needs a corresponding classifier model is then trained on normal and watermarked texts to detect the statistical signature of the watermark. The classifier makes detecting the watermark computationally efficient and does not require access to the underlying LLM. Watermark detectors can be trained with a few thousand examples of normal text and responses that have been watermarked with the specified configuration.
SynthID builds uses a novel sampling algorithm called “Tournament sampling,” a multi-stage process that chooses the next token when creating watermarks.
DeepMind researchers conducted a live experiment with SynthID on nearly 20 million responses generated by Gemini models. Their findings show that SynthID was able to preserve response qualities while also remaining detectable by their classifiers.
Like other watermarking techniques, SynthID has limitations. For example, if a query requires specific factual responses, using SynthID might reduce the quality of the output. The accuracy of the watermark detector can also drop considerably when the text is rewritten thoroughly, though it is resilient to minor edits.
In collaboration with Hugging Face, DeepMind has integrated SynthID into the popular Transformers library, making it easy for developers to add watermarking features to existing applications.
Read more about SynthID on VentureBeat
Read the technical paper of SynthID in Nature
Read more about SynthID Text implementation on the Hugging Face blog