Privacy-preserving AI gets a boost with Google’s VaultGemma model

Differential privacy comes to large language models at scale with Google’s newly launched VaultGemma.
By enabling private AI training at scale, Google's VaultGemma demonstrates that security need not come at the cost of speed.

Google has unveiled VaultGemma, a new large language model built to offer cutting-edge privacy through differential privacy. The 1-billion-parameter model is based on Google’s Gemma architecture and is described as the most powerful differentially private LLM to date.

Differential privacy adds mathematical noise to data, preventing the identification of individuals while still producing accurate overall results. The method has long been used in regulated industries, but has been challenging to apply to large language models without compromising performance.

VaultGemma is designed to eliminate that trade-off. Google states that the model can be trained and deployed with differential privacy enabled, while maintaining comparable stability and efficiency to non-private LLMs.

This breakthrough could have significant implications for developers building privacy-sensitive AI systems, ranging from healthcare and finance to government services. It demonstrates that sensitive data can be protected without sacrificing speed or accuracy.

Google’s research teams say the model will be released with open-source tools to help others adopt privacy-preserving techniques. The move comes amid rising regulatory and public scrutiny over how AI systems handle personal data.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!