Bridging the gap between debiasing and privacy for deep learning

Abstract

The broad availability of computational resources and the recent scientific progresses made deep learning the elected class of algorithms to solve complex tasks. Besides their deployment, two problems have risen: fighting biases in data and privacy preservation of sensitive attributes. Many solutions have been proposed, some of which deepen their roots in the pre-deep learning theory. There are many similarities between debiasing and privacy preserving approaches: how far apart are these two worlds, when the private information overlaps the bias?In this work we investigate the possibility of deploying debiasing strategies also to prevent privacy leakage. In particular, empirically testing on state-of-the-art datasets, we observe that there exists a subset of debiasing approaches which are also suitable for privacy preservation. We identify as the discrimen the capability of effectively hiding the biased information, rather than simply re-weighting it.