Towards privacy-preserving classification in neural networks
Mehmood Baryalai, Julian Jang‐Jaccard, Dongxi Liu · 2016
The requirement for data privacy is limiting to exploit the full potential of what modern data analytic capability could offer. To address such privacy concern, a number of techniques based on homomorphic encryption (HE) have been proposed to allow analytic computation, such as classification based on machine learning techniques, to run on encrypted data. However, these HE-based techniques suffer from a heavy computation overhead due to cryptographic computations having to be done on the encrypted data. We propose a non-colluding dual cloud system that utilizes Paillier cryptosystem. We illustrate how our proposal could reduce inherent computation overhead many similar techniques suffer. Such reduction could make our proposed system to be an ideal solution to use in the real world application.