Transforming Data through Deidentification

Text data shared by users is important for improving the user experience and quality of products and introducing new features. To clarify, text data does not relate to private communications data but rather refers to data that can come from various features e.g., search queries, text captions from public content. Using text data for downstream tasks like analysis, improving user safety, training models can pose significant privacy risks as the data may contain sensitive and private information about individuals. To address these challenges, we introduced a novel text deidentification workflow, designed to improve privacy while maximizing the utility for downstream tasks. The deidentification workflow works as follows: firstly, a PII redaction process systematically eliminates user-identifying attributes; secondly, an LLM rewrite modifies sentences to remove user-specific writing styles and lastly, a validation process gauges the efficacy of text deidentification. In this talk, we will go over the details of the various phases of the text deidentification workflow along with an overview of its implementation using Temporal.