IPCL Tutorial¶

Intel Paillier Cryptosystem Library¶

Intel Paillier Cryptosystem Library (IPCL) is an open-source library which provides accelerated performance of a partial homomorphic encryption (HE), named Paillier cryptosystem, by utilizing Intel® Integrated Performance Primitives Cryptography technologies on Intel CPUs supporting the AVX512IFMA instructions. The library is written in modern standard C++ and provides the essential API for the Paillier cryptosystem scheme. Intel Paillier Cryptosystem Library is certified for ISO compliance.

Enable IPCL as Paillier Scheme in FATE¶

Intel Paillier Cryptosystem Library - Python (ipcl_python), a Python extension package of IPCL, is enabled in FATE as an optional Paillier scheme, alongside original FATE Paillier scheme.

The following sections introduce:

How to enable IPCL when building FATE
How to use IPCL as Paillier scheme when performing privacy-preseving machine learning tasks
How IPCL is integrated into FATE and corresponding code changes

Build FATE with IPCL¶

Build Docker Images

To build FATE images with ipcl_python enabled, please refer to README of FATE-Builder. Set Build_IPCL=1.

IPCL Usage¶

In machine learning task

Set the value of encrypt_param.method to "ipcl" when configuring the model in federatedml.components.

In fate_test
Install fate_test according to fate_test_tutorial
Test Paillier operations with both original FATE Paillier scheme and IPCL
```
fate_test op-test paillier
```

Integration Steps¶

The following steps briefly introduce how IPCL is integrated into FATE, which have been already done since FATE v1.10.0.

Add IPCL as an encryption parameter consts.PAILLIER_IPCL

python/federatedml/util/consts.py

  PAILLIER = 'Paillier'
+ PAILLIER_IPCL = 'IPCL'

python/federatedml/param/encrypt_param.py

+ elif user_input == "ipcl":
+     self.method = consts.PAILLIER_IPCL

Add ipcl_python as the backend of secure protocol

python/federatedml/secureprotocol/encrypt.py

+ from ipcl_python import PaillierKeypair as IpclPaillierKeypair
+ from ipcl_python import PaillierEncryptedNumber as IpclPaillierEncryptedNumber
  ...
+ class IpclPaillierEncrypt(Encrypt):
+ ...

python/federatedml/secureprotocol/__init__.py

- from federatedml.secureprotol.encrypt import RsaEncrypt, PaillierEncrypt
+ from federatedml.secureprotol.encrypt import RsaEncrypt, PaillierEncrypt, IpclPaillierEncrypt
  from federatedml.secureprotol.encrypt_mode import EncryptModeCalculator

- __all__ = ['RsaEncrypt', 'PaillierEncrypt', 'EncryptModeCalculator']
+ __all__ = ['RsaEncrypt', 'PaillierEncrypt', 'IpclPaillierEncrypt', 'EncryptModeCalculator']

Add IpclPaillierEncrypt as the model's cipher_operator, e.g., hetero_logistic_regression

python/federatedml/linear_model/coordinated_linear_model/logistic_regression/hetero_logistic_regression/hetero_lr_base.py

- from federatedml.secureprotol import PaillierEncrypt
+ from federatedml.secureprotol import PaillierEncrypt, IpclPaillierEncrypt
  ...
-         self.cipher_operator = PaillierEncrypt()
+
+         if params.encrypt_param.method == consts.PAILLIER:
+             self.cipher_operator = PaillierEncrypt()
+         elif params.encrypt_param.method == consts.PAILLIER_IPCL:
+             self.cipher_operator = IpclPaillierEncrypt()
+         else:
+             raise ValueError(f"Unsupported encryption method: {params.encrypt_param.method}")
+

Current Status¶

For now, hetero_logistic_regression and hetero_secure_boost models are supported by IPCL.

最后更新: 2022-11-25