search

intel / pailliercryptolib_python

Intel Paillier Cryptosystem Library is an open-source library which provides accelerated performance of a partial homomorphic encryption (HE), named Paillier cryptosystem, by utilizing Intel® IPP-Crypto 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 - Python is a Python extension package intended for Python based privacy preserving machine learning solutions which utilizes the partial HE scheme for increased data and model protection.
Apache License 2.0
56 stars 12 forks source link

Maximum exponent limitation #33

Closed fangxiaoran closed 1 year ago

fangxiaoran commented 1 year ago

It seems sys.float_info.mant_dig is not a proper limitation for maximum exponent.

https://github.com/intel/pailliercryptolib_python/blob/9000f1f5048b66397ab5e61e05ed2ff666158eba/src/ipcl_python/bindings/fixedpoint.py#L29

fangxiaoran commented 1 year ago

According to Wikipedia, a float number is represented as shown below.

image

Thus, multiplying a float number by $2^{(52-e+1023)}$ is enough to keep all significands when converting it to an integer. It's reasonable to use sys.float_info.mant_dig. Note that flt_exponent = math.frexp(scalar)[1] equals to $e-1023+1$, cls.FLOAT_MANTISSA_BITS - flt_exponent $=53-(e-1023+1)=52-e+1023$.

However, a precision loss may occur in math.floor(lsb_exponent / cls.LOG2_BASE), since 16 is chosen as the BASE, not 2. That might be another issue.