Privacy in distributed machine learning

This research explores ways to effectively use distributed machine learning while preserving
\nprivacy. Distributed learning was done on a client-server architecture where each client is an
\nindividual learner training on his or her own dataset and the server exchanges gradients between
\nlearners. Each learner used a convolutional neural network to learn the MNIST dataset
\n(handwritten digits). Only gradients are exchanged instead of training data resulting in some
\nprivacy, but additional privacy was added by multiplying random weights to these gradients.
\nThe first step of the research was to replicate results from Shokri and Shmatikov’s 2015 paper
\non privacy-preserving deep Learning. Shokri and Shmatikov’s paper primarily worked with
\nmultiple clients and a single server. The research extends this architecture by creating a
\nmultiple server and multiple client architecture and adding random weights to the gradients.
\nThe research successfully produced empirical results showcasing the effectiveness of our
\ndistributed machine learning algorithm since it was able to maintain an acceptable classification
\naccuracy while providing a certain degree of privacy.
\nIn addition to Shokri and Shmatikov’s paper, this research draws upon Gade and Vaidya’s 2016
\npaper “Distributed Optimization for Client-Server Architecture with Negative Gradient
\nWeights”.