Application Testing of Generative Adversarial Privacy

To protect personal privacy from inference attacks, using techniques from Generative Adversarial Networks (GANs), we have implemented a simple Generative Adversarial Privacy Architecture (GAP) architecture composed of an encoders, a distortion metric, and two classifiers to illustrate the concept. This architecture distorts input data to hinder the predictions of one classifier while minimally affecting the accuracy of the other classifier. This is directly related to deterring an inference attack by lowering the ability to infer sensitive private information while allowing prediction of nonsensitive public information.