Predicting the Performance of Privacy-Preserving Data Analytics Using Architecture Modelling and Simulation

Privacy-preserving data analytics is an emerging technology which allows multiple parties to perform joint data analytics without disclosing source data to each other or a trusted third-party. A variety of platforms and protocols have been proposed in this domain. However, these systems are not yet widely used, and little is known about them from a software architecture and performance perspective. Here we investigate the feasibility of using architectural performance modelling and simulation tools for predicting the performance of privacy-preserving data analytics systems. We report on a lab-based experimental study of a privacy-preserving credit scoring application that uses an implementation of a partial homomorphic encryption scheme. The main experiments are on the impact of analytic problem size (number of data items and number of features), and cryptographic key length for the overall system performance. Our modelling approach performed with a relative error consistently under 5\% when predicting the median learning time for the scoring application. We find that the use of this approach is feasible in this technology domain, and discuss how it can support architectural decision making on trade-offs between properties such as performance, cost, and security. We expect this to enable the evaluation and optimisation of privacy-preserving data analytics technologies.