Hugging Face has introduced a new feature that enables developers to run privacy-preserving inferences on its endpoints, allowing sensitive data to be processed without exposing it to the cloud service. This capability leverages confidential computing techniques to encrypt data during processing, ensuring that even the AI model provider cannot access the raw inputs or outputs.
The solution uses Intel SGX enclaves, which create a secure area in the server's memory where data is decrypted and processed. This means that user queries are protected from the moment they leave the user's device until the inference result is returned. Developers can deploy existing models from the Hugging Face Hub or their own custom models onto these secure endpoints.
This development is particularly important for industries like healthcare, finance, and legal services, where data privacy regulations are strict. By running inferences in a hardware-based trusted execution environment, organizations can leverage powerful AI models without compromising on compliance.
Hugging Face plans to expand the feature to support more hardware and cloud providers in the future, making privacy-preserving AI more accessible.