Google izlaiž atvērtā koda AI modeli VaultGemma ar diferenciālu privātumu

Google has announced a new open model called VaultGemma, based on recent findings in differential privacy. The AI ​​model has 1 billion parameters and is based on Gemma 2, part of Google’s family of small language models.

Historically, it has been known that adding differential privacy to an AI model could prevent it from regurgitating an output that is identical to the data it was trained on. This is done by introducing a small amount of noise during training.

However, adding noise to training data reduces the accuracy of an AI model.

Google collaborated with Google DeepMind on a study called “Scaling Differentially Private Language Models,” which explored the amount of randomized noise compared to the total training dataset.

The researchers conducted experiments with different model sizes and noise-to-batch ratios to understand how differential privacy could be reduced while maintaining good power.

A blog post by Google claimed that this is the largest open model for differential privacy that can be used to develop high-usage AI models.

Developers can download VaultGemma from Face and Kaggle Hugs. Google has also released scales so users can fine-tune the AI ​​model to create their own versions.