Presto-Native Noisy Aggregations for Privacy-Preserving Workflows

At Meta, large-scale data analysis happens constantly, across varied surfaces, platforms, and systems. Differential privacy (DP), because of its strong protection, is one of the privacy-enhancing technologies deployed by Meta to protect users' privacy. However, implementing DP in practice, especially at Meta scale, has many challenges, including the diversity of interfaces for analysis, size of datasets, expertise required, and integration with other policy requirements and enforcement. In this talk, we describe an approach to private data analysis at Meta that places a set of common privacy primitives in the compute engine (Presto), which are leveraged by different frameworks and services to enforce DP guarantees across our many systems. Examples include automatic query rewriting for interactive data analysis, privacy-preserving ETL pipelines, and web mapping of aggregate statistics. The Presto-based approach helped increase flexibility, minimize changes to existing workflows, and enable robust privacy enforcement and guarantees. This is joint work with Jonathan Hehir (Meta Platforms, Inc.)