DP-Rewrite: Towards Reproducibility and Transparency in Differentially Private Text Rewriting

Proceedings of the 29th International Conference on Computational Linguistics, pages 2927–2933 October 12–17, 2022. 2927 DP-Rewrite: Towards Reproducibility and Transparency in Differentially Private Text Rewriting Timour Igamberdiev1 and Thomas Arnold2 and Ivan Habernal1 1Trustworthy Human Language Technologies 2Ubiquitous Knowledge Processing Lab Department of Computer Science, Technical University of Darmstadt {timour.igamberdiev,ivan.habernal}@tu-darmstadt.de arnold@ukp.informatik.tu-darmstadt.de www.trusthlt.org www.ukp.tu-darmstadt.de Abstract Text rewriting with differential privacy (DP) provides concrete theoretical guarantees for protecting the privacy of individuals in tex- tual documents. In practice, existing systems may lack the means to validate their privacy- preserving claims, leading to problems of trans- parency and reproducibility. We introduce DP-Rewrite, an open-source framework for differentially private text rewriting which aims to solve these problems by being modular, ex- tensible, and highly customizable. Our system incorporates a variety of downstream datasets, models, pre-training procedures, and evalua- tion metrics to provide a flexible way to lead and validate private text rewriting research. To demonstrate our software in practice, we pro- vide a set of experiments as a case study on the ADePT DP text rewriting system, detecting a privacy leak in its pre-training approach. Our system is publicly available, and we hope that it will help the community to make DP text rewriting research more accessible and trans- parent. 1 Introduction Protecting the privacy of individuals has been gaining attention in NLP. One particular setup is text rewriting using local differential privacy (DP) (Dwork and Roth, 2013), which provides proba- bilistic guarantees of ‘how much’ privacy can be lost in the worst case if an individual gives us their piece of text that has been rewritten with DP. For instance, given a text “I want to fly from Newark to Cleveland on Friday”, the system may rewrite it as “Flights from Los Angeles to Houston this week”. Only a few recent works have touched on this challenging topic. For example, Krishna et al. (2021) proposed ADePT: A text rewriting system based on the Laplace mechanism. However, it turned out that their DP method was formally flawed (Habernal, 2021). We also see another recent approach, DP-VAE (Weggenmann et al., 2022), which shows results that look surprisingly good for the level of guaranteed privacy. How- ever, neither ADePT nor DP-VAE published their source codes, so the community has no means to perform any empirical checks to validate the privacy-preserving claims. Therefore, the lack of transparency and reproducibility is the main ob- stacle to the accountability of DP text-rewriting systems. We asked whether an open, modular, easily ex- tensible, and highly customizable framework for differentially private text rewriting could help the community gain further insight into the utility and potential pitfalls of such systems. We hypothesize that by integrating various downstream datasets, models, pre-training procedures, and evaluation metrics into one software package, we improve the transparency, accountability, and reproducibility of research in differentially private text rewriting. Our main contributions are twofold. First, this demo paper presents DP-Rewrite, an open- source framework for differentially private text rewriting experiments. It includes a correct reimple- mentation of ADePT as …

Content truncated to 500 words