Google has launched VaultGemma, a new artificial intelligence (AI) model designed to protect sensitive information without losing performance. The one-billion-parameter model uses differential privacy, a technology that prevents AI from remembering or exposing private data.
The model was created by Google Research and Google DeepMind and is built on Google’s Gemma architecture. According to Google, VaultGemma proves that AI can be trained to keep information private without becoming less powerful.
Differential privacy works by adding small amounts of “noise” to training data. This makes it impossible to trace or extract specific details while still allowing the AI to function normally. VaultGemma was designed and trained with this system from the start, ensuring it cannot leak sensitive data.
Google also announced new “DP Scaling Laws”, which make it possible to train larger private AI models. Normally, adding noise makes training unstable and expensive. However, Google’s researchers say they developed new methods to handle these challenges and reduce costs, making the technology more practical.
Tests show that VaultGemma performs almost the same as non-private Gemma models on key benchmarks such as MMLU and Big-Bench. This is a major step forward because older private AI models often performed much worse.
VaultGemma is based on a decoder-only transformer design with 26 layers and Multi-Query Attention. It processes sequences of up to 1,024 tokens, keeping the heavy demands of private training manageable.
In a move toward openness, Google is making VaultGemma available as open source on Hugging Face and Kaggle. This is different from proprietary AI tools such as Gemini Pro. The company hopes the technology will be useful in fields like healthcare, where AI can analyze patient data without putting privacy at risk.
By keeping its training data secret, VaultGemma also reduces risks linked to misinformation and bias, according to Google researchers.
 model designed to protect sensitive information){kind=link}