Amazon SageMaker AI + FHE enables encrypted ML inference without decryption

Amazon SageMaker AI now integrates with fully homomorphic encryption (FHE) via the concrete-ml library, allowing ML inference to be performed on encrypted data without ever decrypting it. The blog post walks through training a concrete-ml model (compatible with scikit-learn) in a custom SageMaker container, deploying it to an endpoint, and creating a client that encrypts queries before sending them. The model processes the encrypted input and returns an encrypted prediction, which the client decrypts locally. This ensures that the cloud provider—including SageMaker itself—never sees the raw data.

Use cases span regulated industries: healthcare (predicting medical outcomes without exposing patient data), energy (analyzing satellite images of sensitive sites), and telecom (detecting spam without reading customer messages). The approach complements AWS Nitro Enclaves: while Nitro relies on hardware isolation, FHE provides mathematical security where data stays encrypted end-to-end. The solution requires data normalization to a uniform scale (e.g., [-1, 1]) and supports several model types out of the box. This marks a significant step toward practical privacy-preserving machine learning in the cloud.