Concrete ML

Converts trained scikit-learn and PyTorch models into FHE-compatible circuits with minimal code changes. The framework analyzes the model architecture and automatically replaces non-linear operations with polynomial approximations suitable for homomorphic execution.

• Supports linear models, tree-based models, and neural networks
• Handles quantization-aware training internally
• Preserves model accuracy within configurable precision bounds
• Single API call: compile(model, data) to produce an FHE circuit

Built on the TFHE scheme, Concrete ML leverages programmable bootstrapping to evaluate non-linear activation functions directly on encrypted data. This enables efficient gate-by-gate computation with fast bootstrapping cycles.

• Supports arbitrary boolean and integer operations on ciphertexts
• Programmable bootstrapping evaluates lookup tables during noise refresh
• Eliminates the need for deep polynomial approximations of activation functions
• Optimized for low-latency single-inference scenarios

Automatically quantizes floating-point model weights and activations to low-precision integers required by FHE circuits. The framework performs quantization-aware training to minimize accuracy degradation when converting continuous values to discrete ciphertext-compatible representations.

• Configurable bit-width: typically 2-8 bits per weight
• Post-training quantization and quantization-aware training modes
• Automatic calibration using representative input data
• Balances precision against ciphertext noise budget consumption

Implements a client-server architecture where the client encrypts input data, the server performs FHE inference on the encrypted model, and only the client can decrypt the result. The server never sees raw inputs or outputs.

• Client holds the secret decryption key
• Server holds the compiled FHE circuit and evaluation keys
• Encrypted predictions returned to client for local decryption
• Suitable for privacy-preserving ML-as-a-service deployments

Provides drop-in replacements for popular scikit-learn estimators and PyTorch modules. Developers can train models using familiar APIs and convert them to FHE circuits without rewriting model logic.

• Supported scikit-learn models: LinearRegression, LogisticRegression, DecisionTree, RandomForest, XGBoost
• Supported PyTorch layers: Linear, Conv2d, ReLU, Sigmoid, MaxPool2d
• Custom neural network architectures via torch.nn.Module subclassing
• ONNX import pipeline for broader model format support

Includes a simulation mode that executes the FHE circuit logic in plaintext, enabling rapid development and debugging without the computational overhead of actual homomorphic encryption. Identical API, identical results, but orders of magnitude faster.

• Bit-exact results matching real FHE execution
• Enables hyperparameter tuning and accuracy validation
• Detects compilation errors before expensive FHE execution
• Seamless toggle between simulation and FHE modes

The foundational Fully Homomorphic Encryption scheme optimized for fast gate-by-gate bootstrapping of binary circuits. Unlike leveled schemes, TFHE enables efficient evaluation of arbitrary boolean functions on encrypted bits, making it ideal for the decision trees and quantized neural networks that Concrete ML converts. Its programmable bootstrapping extension is the key mechanism that evaluates non-linear activation functions during encrypted inference.

A community-driven open-source FHE library providing a unified API for multiple schemes including BFV, BGV, CKKS, and TFHE. While Concrete ML is built specifically on Zama's TFHE implementation, OpenFHE represents the broader ecosystem of interoperable FHE tooling. It supports hardware acceleration and multi-key extensions, making it a common choice for research teams exploring hybrid privacy-preserving computation architectures.