Utility-Preserving Anonymization of Textual Documents

Identifier:  TDX:3182

Handle: https://hdl.handle.net/20.500.11797/TDX3182

Authors:  Hassan, FadiAbdulfattah Mohammed

Abstract:
Every day, people post a significant amount of data on the Internet, such as tweets, reviews, photos, and videos. Organizations collecting these types of data use them to extract information in order to improve their services or for commercial purposes. Yet, if the collected data contain sensitive personal information, they cannot be shared with third parties or released publicly without consent or adequate protection of the data subjects. Privacy-preserving mechanisms provide ways to sanitize data so that identities and/or confidential attributes are not disclosed. A great variety of mechanisms have been proposed to anonymize structured databases with numerical and categorical attributes; however, automatically protecting unstructured textual data has received much less attention. In general, textual data anonymization requires, first, to detect pieces of text that may disclose sensitive information and, then, to mask those pieces via suppression or generalization. In this work, we leverage several technologies to anonymize textual documents. We first improve state-of-the-art techniques based on sequence labeling. After that, we extend them to make them more aligned with the notion of privacy risk and the privacy requirements. Finally, we propose a complete framework based on word embedding models that captures a broader notion of data protection and provides flexible protection driven by privacy requirements. We also leverage ontologies to preserve the utility of the masked text, that is, its semantics and readability. Extensive experimental results show that our methods outperform the state of the art by providing more robust anonymization while reasonably preserving the utility of the protected outcomes