Deep Neural Networks as Similitude Models for Sharing Big Data
Philip Derbeko, Shlomi Dolev, Ehud Gudes · 2019
The amount of data grows rapidly with time and shows no signs of stopping. Ubiquitous computing continues to collect and generate more and more data as both the number of devices grows and the capabilities of devices increase. We suggest processing the data on end devices by building a representative model of the data (“similitude” model). Sharing a smaller model instead of the entire data allows for saving computing power, network time, processing time and also, keeping the collected data private. In the past research, we suggested the use of similitude models, as compact models of data representation instead of the data itself. In this paper, we suggest the use of deep neural networks (DNN) as a data model to answer different types of queries. More specifically, we show that by building two models (generative network and auto-encoder) it is possible to answer approximately both statistical queries and membership queries without exposing the entire dataset.