Building Decentralized Image Classifiers with Federated Learning

The commercial use of neural networks has been greatly curbed by data privacy concerns. As long as the accumulation and use of private data is regarded necessary for integrating neural networks into products, consumers will be reluctant to use or allow access to any deep learning integrated product and producers will be equally deterred from leveraging deep learning for performance improvement. Federated learning was first introduced as a solution to this conundrum in a 2016 paper published by Google titled Communication-Efficient Learning of Deep Networks from Decentralized Data [1]. In this study, we examine how the performance of a decentralized image classifier compares to that of a centralized one. The performance of an image classifier trained across ten devices was compared to a model built with the same architecture but trained centrally on one corpus of training data. The outcome demonstrates that the decentralized model compares quite well to the centrally trained classifier in terms of accuracy, precision and recall.