Google has unveiled VaultGemma, a one-billion-parameter privacy-focused AI model trained using differential privacy techniques to prevent memorisation of sensitive data. By adding calibrated noise during training, the model avoids exposing personal identifiers like addresses or credit cards. While this approach reduces accuracy and increases compute costs, it shows privacy-preserving AI is feasible, potentially setting a new industry standard for safeguarding user data.

About The Author

Bharat Journal

See author's posts