Towards Task-Agnostic Privacy- and Utility-Preserving Models

Proceedings of Recent Advances in Natural Language Processing, pages 394–401 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_045 394 Towards Task-Agnostic Privacy- And Utility-Preserving Models Yaroslav Emelyanov Sberbank, Moscow, Russia emelyanov.y.i@sberbank.ru Abstract Modern deep learning models for natural language processing rely heavily on large amounts of annotated texts. However, obtain- ing such texts may be difficult when they con- tain personal or confidential information, for example, in health or legal domains. In this work, we propose a method of de-identifying free-form text documents by carefully redact- ing sensitive data in them. We show that our method preserves data utility for text classifica- tion, sequence labeling, and question answer- ing tasks. 1 Introduction Data privacy has become an important topic re- cently, and new regulations that govern the pro- cessing and usage of consumer personal data arise every year. We use the term sensitive data to de- fine all information that contains personal data or other potentially compromising data that can lead to the re-identification of individuals or organiza- tions. With the advent of machine learning, privacy and compliant data governance have become even more important. Free-form text documents often contain sensitive data. For example, legal documents contain full information about individuals or organizations, di- alogues contain references to different people and possibly information that can lead to their identi- fication, such as addresses or job positions. NLP applications such as text classification or question answering require annotated data, and as Feyisetan et al. (2019) argues, the privacy cost of annotat- ing texts must be considered when developing new applications. Anonymization, or de-identification, is viewed as one of the ways to make working with non- public datasets both safer and more compliant. It is also important to preserve utility in de-identified data for further usages, e.g. machine learning or analytics. Traditional approaches like k-anonymity (Sweeney, 2002) were shown to be ineffective when applied to high-dimensional media (Aggar- wal, 2005) for example, free-form texts. More- over, unlike structured data present in relational databases, anonymization of unstructured text doc- uments poses additional challenges, because loca- tions of textual spans which constitute sensitive data must be inferred at runtime. Previous work either focused on preventing the model from memorizing data (Kerrigan et al., 2020) or considered only the text classification task. In this paper, we aim to solve the challenges of pre- serving both privacy and utility in free-form text documents with respect to different NLP tasks. We propose a simple “find-and-replace” method for au- tomatically de-identifying documents and evaluate the utility of the de-identified data with respect to different downstream tasks, namely named entity recognition, question answering and text classifica- tion. To the best of our knowledge, this is the first work that evaluates the utility of the de-identified data for sequence labeling and reading comprehen- sion tasks. Specifically, our contributions are as follows: 1. We develop and evaluate the deep-learning- based method for de-identifying text docu- ments. 2. We conduct series of experiments and show that our method mostly preserves utility for different NLP tasks: NER, …

Content truncated to 500 words