De-Identification of Privacy Sensitive Information in Resumes with GPT-4: An Utility Analysis for Automated Job Role Classification

Abstract

As organizations face the challenge of managing large amounts of data, privacy concerns have become increasingly prevalent when sharing sensitive privacy information with machine learning experts. This paper addresses the fundamental issue of privacy-sensitive information de-identification by introducing in-prompt de-identification, an approach that exploits the capabilities of large language models. Existing de-identification techniques often struggle to ensure complete privacy, and methods with higher privacy often result in a loss of data utility. In contrast, in-prompt de-identification is capable of generating synthetic, human-readable data samples from given inputs and bridges the gap between privacy and utility. With this article, we contribute to the de-identification of real-world resume data using in-prompt de-identification based on OpenAI’s GPT-4. Notably, our classification model, trained on GPT-4 generated data, shows no significant loss in performance compared to our baseline model trained on the original data.