Adversarial representation learning for synthetic replacement of private attributes

Data privacy is an increasingly important aspect of many real-world Data\nsources that contain sensitive information may have immense potential which\ncould be unlocked using the right privacy enhancing transformations, but\ncurrent methods often fail to produce convincing output. Furthermore, finding\nthe right balance between privacy and utility is often a tricky trade-off. In\nthis work, we propose a novel approach for data privatization, which involves\ntwo steps: in the first step, it removes the sensitive information, and in the\nsecond step, it replaces this information with an independent random sample.\nOur method builds on adversarial representation learning which ensures strong\nprivacy by training the model to fool an increasingly strong adversary. While\nprevious methods only aim at obfuscating the sensitive information, we find\nthat adding new random information in its place strengthens the provided\nprivacy and provides better utility at any given level of privacy. The result\nis an approach that can provide stronger privatization on image data, and yet\nbe preserving both the domain and the utility of the inputs, entirely\nindependent of the downstream task.\n