Google DeepMind 推出 VaultGemma：全球最強大的差分隱私（Differential Privacy）開源大語言模型

Google DeepMind has officially launched VaultGemma, currently the world's most capable large language model (LLM) trained entirely from scratch using Differential Privacy (DP) technology. In the age of artificial intelligence, preventing models from "memorizing" sensitive training data during the training process — and inadvertently leaking it later (such as personally identifiable information, medical records, or confidential business documents) — has been a central challenge in the field of privacy protection. VaultGemma's arrival provides a mathematically rigorous solution to this challenge.