Benchmarking Privacy Preserving Scientific Operations

Abstract

In this work, we examine the efficiency of protocols for secure evaluation of basic mathematical functions (sqrt, sin, arcsin, amongst others), essential to various application domains. e.g., Artificial Intelligence. Furthermore, we have incorporated our code in state-of-the-art Multiparty Computation (MPC) software, so we can focus on the algorithms to be used as opposed to the underlying MPC system. We make use of practical approaches that, although, some of them, theoretically can be regarded as less efficient, can, nonetheless, be implemented in such software libraries without further adaptation. We focus on basic scientific operations, and introduce a series of data-oblivious protocols based on fixed point representation techniques. Our protocols do not reveal intermediate values and do not need special adaptations from the underlying MPC protocols. We include extensive computational experimentation under various settings and MPC protocols.

Secure Quantized Training for Deep Learning

“…Logarithm Computation logarithm with any public base can be reduced to logarithm to base two using log x y = log 2 y • log x 2. Aly and Smart [2019] propose to represent y as y = a • 2 b , where a ∈ [0.5, 1) and b ∈ Z. This then allows to compute log 2 y = log 2 a + b.…”

“…Exponentiation By using x y = 2 y log 2 x , any exponentiation can be reduced to exponentiation with base two. Aly and Smart [2019] show how to compute 2 a = 2 a • 2 a− a by computing the two exponents using bit decomposition and the second factor using a polynomial approximation. Regarding the first factor, if…”

“…Inverse square root Aly and Smart [2019] propose to compute square root using Goldschmidt and Raphson-Newton iterations. We could combine this with the division operator introduced above.…”

A Note on Our Submission to Track 4 of iDASH 2019

“…Standard logistic regression uses the sigmoid function based on the exponential, and computing the loss the logarithm function. Our implementation of these is based on the code provided in SCALE-MAMBA [4] by Aly and Smart [1].…”

“…In this section, we summarize the task. 1 Participants were invited to implement machine learning (ML) training in multi-party computation (MPC). MPC allows a set of parties to jointly compute on data hold among each other without revealing anything other than the result of the computation.…”