Privacy-preserving Representation Learning by Disentanglement

Tassilo Klein, Moin Nabi

Abstract: Deep learning and latest machine learning technology heralded an era of success in data analysis. Accompanied by the ever increasing performance, reaching super-human performance in many areas, is the requirement of amassing more and more data to train these models. Often ignored or underestimated, the big data curation is associated with the risk of privacy leakages. The proposed approach seeks to mitigate these privacy issues. In order to sanitize data from sensitive content, we propose to learn a privacy-preserving data representation by disentangling into public and private part, with the public part being shareable without privacy infringement. The proposed approach deals with the setting where the private features are not explicit, and is estimated though the course of learning. This is particularly appealing, when the notion of sensitive attribute is ``fuzzy''. We showcase feasibility in terms of classification of facial attributes and identity on the CelebA dataset. The results suggest that private component can be removed in the cases where the the downstream task is known a priori (i.e., ``supervised''), and the case where it is not known a priori (i.e., ``weakly-supervised'').