SwaNN: Switching among Cryptographic Tools for Privacy-preserving Neural Network Predictions

Gamze Tillem, Beyza Bozdemir, Melek Önen

Abstract

The rise of cloud computing technology led to a paradigm shift in technological services that enabled enterprises to delegate their data analytics tasks to cloud servers which have domain-specific expertise and computational resources for the required analytics. Machine Learning as a Service (MLaaS) is one such service which provides the enterprises to perform machine learning tasks on the cloud. Despite the advantage of eliminating the need for computational resources and domain expertise, sharing sensitive data with the cloud server brings a privacy risk to the enterprises. In this paper, we propose SwaNN, a protocol to privately perform neural network predictions for MLaaS. SwaNN brings together two well-known techniques for secure computation: partially homomorphic encryption and secure two-party computation, and computes neural network predictions by switching between the two methods. The hybrid nature of SwaNN enables to maintain the accuracy of predictions and to optimize the computation time and bandwidth usage. Our experiments show that SwaNN achieves a good balance between computation and communication cost in neural network predictions compared to the state-of-the-art proposals.

Download


Paper Citation