Lilly Tech Systems
HE Libraries & Tools
A practical guide to the open-source homomorphic encryption libraries available for building encrypted ML applications. From low-level crypto libraries to high-level ML frameworks.
Library Comparison
| Library | Maintainer | Schemes | Level | Best For |
|---|---|---|---|---|
| Microsoft SEAL | Microsoft | BFV, BGV, CKKS | Low-level | Custom HE applications |
| Concrete ML | Zama | TFHE | High-level | ML model compilation to FHE |
| TenSEAL | OpenMined | BFV, CKKS | Mid-level | Tensor operations on encrypted data |
| OpenFHE | Duality Technologies | BFV, BGV, CKKS, TFHE | Low-level | Research, full FHE support |
| HElib | IBM | BGV, CKKS | Low-level | Research, bootstrapping |
Microsoft SEAL
SEAL is the most widely used HE library, providing well-documented implementations of BFV and CKKS:
import seal # Setup CKKS parameters parms = seal.EncryptionParameters(seal.scheme_type.ckks) parms.set_poly_modulus_degree(8192) parms.set_coeff_modulus(seal.CoeffModulus.Create( 8192, [60, 40, 40, 60])) context = seal.SEALContext(parms) keygen = seal.KeyGenerator(context) public_key = keygen.create_public_key() secret_key = keygen.secret_key() encryptor = seal.Encryptor(context, public_key) evaluator = seal.Evaluator(context) decryptor = seal.Decryptor(context, secret_key) encoder = seal.CKKSEncoder(context) # Encode and encrypt scale = 2.0**40 plain = encoder.encode([1.5, 2.3, 3.7], scale) encrypted = encryptor.encrypt(plain) # Compute on encrypted data plain_coeff = encoder.encode([2.0, 2.0, 2.0], scale) evaluator.multiply_plain_inplace(encrypted, plain_coeff) evaluator.rescale_to_next_inplace(encrypted) # Decrypt result result = decryptor.decrypt(encrypted) decoded = encoder.decode(result) # [3.0, 4.6, 7.4]
Concrete ML
Concrete ML by Zama is the highest-level FHE ML library. It can compile standard scikit-learn and PyTorch models to run on encrypted data:
from concrete.ml.sklearn import RandomForestClassifier from sklearn.datasets import load_iris from sklearn.model_selection import train_test_split # Load data and train X, y = load_iris(return_X_y=True) X_train, X_test, y_train, y_test = train_test_split(X, y) model = RandomForestClassifier(n_bits=6, n_estimators=10) model.fit(X_train, y_train) # Compile to FHE circuit fhe_circuit = model.compile(X_train) # Simulate FHE execution (faster for testing) y_pred_simulated = model.predict(X_test, fhe="simulate") # Run actual FHE inference fhe_circuit.client.keygen() y_pred_fhe = model.predict(X_test, fhe="execute") print(f"FHE accuracy: {(y_pred_fhe == y_test).mean():.2%}")
TenSEAL
TenSEAL provides a tensor abstraction over SEAL, making it natural for ML applications. It supports encrypted matrix operations and is well-suited for implementing neural network layers.
OpenFHE
OpenFHE (successor to PALISADE) is the most comprehensive FHE library, supporting all major schemes including TFHE with programmable bootstrapping. It is primarily a C++ library with Python bindings.