zama-ai/concrete-ml

What it solves

Concrete ML allows data scientists to perform machine learning on encrypted data without needing deep expertise in cryptography. It solves the problem of processing sensitive information—such as healthcare records, financial data, or biometric identities—while keeping that data encrypted throughout the entire computation process, ensuring privacy and security.

How it works

The library is built on top of the Concrete framework and utilizes Fully Homomorphic Encryption (FHE). It provides a set of tools that automatically convert standard machine learning models into their homomorphic equivalents. It offers built-in models with APIs similar to scikit-learn and XGBoost, and it also supports the conversion of custom PyTorch or Keras/TensorFlow models via ONNX, using quantization-aware training to make them FHE-compatible.

Who it’s for

It is designed for data scientists who want to implement privacy-preserving machine learning (PPML) using familiar frameworks and APIs, as well as organizations in highly regulated industries like healthcare, finance, and government that need to analyze sensitive data securely.

Highlights

• FHE-friendly built-in models: Ready-to-use models with interfaces equivalent to scikit-learn and XGBoost.
• Custom model support: Ability to import PyTorch or Keras/TensorFlow models via ONNX.
• Privacy-preserving inference: Run predictions on encrypted inputs without decrypting them.
• Broad use cases: Applications ranging from encrypted sentiment analysis and health diagnosis to credit card approval and encrypted image filtering.