Project Lighthouse — Part 3: Introducing project-lighthouse-anonymize
Project Lighthouse, launched in 2020 by Airbnb in collaboration with civil rights and privacy organizations, aims to measure and mitigate potential disparities in user experiences based on perceived race data, which is never linked to individual accounts. This data is only used for disparity measurement, and users can opt-out. Airbnb has open-sourced "project-lighthouse-anonymize," a Python library that powers the project's privacy-preserving anonymization code. Alongside the code release, two new technical papers have been published, complementing the original 2020 paper. The 2020 paper established the privacy-by-design approach, choosing k-anonymity to prevent sensitive attribute disclosure. The first new paper, "Core Mondrian: Basic Mondrian beyond k-anonymity," details a scalable k-anonymity algorithm that extends the classic Mondrian algorithm with an extensible architecture, parallel processing, and other enhancements for large datasets. The second new paper, "Measuring Data Quality for Project Lighthouse," introduces a comprehensive framework for assessing the quality of anonymized data, including three primary metrics: Pearson correlation, Revised Information Loss Metric (RILM), and Normalized Mutual Information v1 (NMIv1). This framework uses an empirical validation methodology, reframing data quality assessment as a machine learning classification problem to predict when anonymized data yields statistically valid results. The library provides default thresholds for these metrics. The library is available on PyPI and GitHub, allowing users to enforce both k-anonymity and p-sensitive k-anonymity via perturbation, with a complete example provided using the UCI Adult dataset. Airbnb emphasizes transparency in its methodologies to build trust and encourage similar initiatives from other companies.