Privately Training an AI Model Using Fake Images Generated by Generative Adversarial Networks

This report was originally published in September 2019.

Abstract

Deep learning algorithms produce sophisticated results using different machine learning and computer vision tasks. To perform well on a given problem, these algorithms require a large dataset for training. Often, deep learning algorithms lack generalization and suffer from over-fitting when trained on a small dataset. For example, the storage of image data along with its corresponding labels for supervised image analysis in medical imaging is costly and time-consuming. Another challenge is that most of the data collected by corporations and public institutions is sensitive and may be prohibited from being shared publicly or with third parties. In this paper, Generative Adversarial Networks (GANs) are used to generate synthetic images that can then be used for further analysis in deep learning algorithms or used by a third party, while obscuring any confidential information. This research has been carried out on proprietary images of race cars using multiple GAN techniques that generate precise segmented images based on car classes. The images generated using one of the techniques (SN-GAN) captured features of real data so well that the classification model trained on those generated images achieved 89.6% accuracy, when tested on real images.