An Adversarial Learning Framework for Privacy Preserving Communications

We develop a machine learning-based approach that allows to achieve privacy in communications by exploiting an advantage at the physical layer. Our goal is to transmit useful data to the intended receiver while preventing sensitive data from leaking to an eavesdropper who has access to the channel. We adopt an adversarial approach involving two competing neural networks to learn efficient coding schemes that allow to regulate the tradeoff between quality and privacy.