Privacy Preserving Deep Learning using Secure Multiparty Computation

Many different types of problems have been solved using deep learning in recent pasts. Deep learning techniques are useful for finding solutions to different types of data type's right from structures to semi structures or unstructured. Problems that are based on clustering, classification regression are effectively implemented using deep learning techniques. This utility of machine and deep learning techniques calls for keeping these services on cloud. Providing machine learning as a service to cloud opens problems of security concern of the data involved in training that belongs to different parties involved in training and also the security concerns arises for the data model being trained. This paper has implemented a privacy preserving technique based on secure multi-party computation that creates secret shared to solve the privacy issues for the data involved in training. Our experimental analysis is carried out using MNIST dataset for hand written character recognition as data for learning problem. Experimental analysis indicated that MNIST dataset can be trained to better accuracy using secure multiparty computation and keep the data secured on the network. The PyTorch and PySyft libraries are used for experimentation.