OpenDP-Open Source Differential Privacy Tools

Overview

OpenDP is a community-driven initiative to build trustworthy, open-source software for differential privacy (DP) — a mathematically rigorous method for extracting statistical insights from sensitive data while providing provable privacy guarantees for the individuals represented in that data. Led by faculty directors Salil Vadhan (Vicky Joseph Professor of Computer Science and Mathematics, Harvard SEAS) and Gary King (Weatherhead University Professor, Harvard), OpenDP is anchored at Harvard University and engages a global community of collaborators from academia, government, and industry. Its core output is the OpenDP Library, available for Python, R, and Rust, which implements vetted, composable differential privacy algorithms that data practitioners can configure and deploy with confidence. OpenDP is funded in part by the Alfred P. Sloan Foundation and by a landmark research collaboration with Microsoft; it receives additional support from other foundations and government agencies.

Background: What Is Differential Privacy

Differential privacy, first formalized by Cynthia Dwork, Frank McSherry, Kobbi Nissim, and Adam Smith in their landmark 2006 paper, is a mathematical framework for releasing statistical information about a dataset while limiting what can be learned about any individual record. It works by adding carefully calibrated random noise to query results, such that the output is nearly indistinguishable whether or not any particular person's data is included. The "privacy loss" incurred by each query is quantified by parameters—typically denoted ε (epsilon)—allowing total privacy expenditure across many queries to be tracked and bounded. Unlike pseudonymization or coarse data suppression, differential privacy offers a formal, composable guarantee: the privacy loss of combining multiple DP releases is bounded by the sum of their individual losses. This property makes it uniquely suited to situations where data is queried repeatedly or shared across institutions. Major adopters of differential privacy include Apple, Google, and Microsoft in their consumer data products, and the US Census Bureau in its 2020 decennial census—the first census in history to publish results protected by differential privacy.

History and Founding

OpenDP grew directly out of Harvard's Privacy Tools Project, an interdisciplinary research initiative established in 2012 through the joint efforts of the Center for Research on Computation and Society, the Berkman Klein Center for Internet and Society, and the Institute for Quantitative Social Science (IQSS). Supported by the National Science Foundation, the Privacy Tools Project advanced theoretical and applied research on data privacy across computer science, statistics, law, and policy. One of its flagship outputs was PSI (a Private data Sharing Interface), an early differentially private data-sharing platform, which demonstrated both the promise and the practical challenges of bringing DP tools to non-expert users.

In 2019, Vadhan and King co-founded OpenDP as a community project to operationalize the Privacy Tools Project's research into production-quality open-source software. A major catalyst was a collaboration with Microsoft, announced publicly in May–June 2020, in which Microsoft contributed engineering expertise, co-developed an open-source DP library (then called SmartNoise), and granted a royalty-free license under its differential privacy patents to the world through the OpenDP initiative. This was, at the time, described as a first-of-its-kind open-source platform for differential privacy. The runtime and validator components were built in Rust for memory safety and performance, with Python bindings for accessibility; R support followed. Microsoft's partnership has remained central to OpenDP's development.

Programs and Tools

OpenDP Library. The core open-source software library, available in Python, R, and Rust, implementing a modular framework of differentially private algorithms and mechanisms. The library is built around composable "measurements" and "transformations" that can be chained into analysis pipelines, with each step contributing to a tracked and bounded privacy loss budget. Version 0.13 (released April 2025) added identifier truncation, synthetic data generation, and linear regression; version 0.14 added odometers and enhancements to core mechanisms. The library is deliberately low-level—prioritizing correctness and rigour over convenience—and is intended as a foundation on which higher-level tools such as SmartNoise SDK are built. All code is open-source on GitHub, with active development contributions from the community.

DP Wizard. A user-friendly graphical interface released in December 2024 (v0.1) and actively developed since, designed to lower the barrier to entry for researchers and data holders who want to use differential privacy without deep expertise in the underlying algorithms. DP Wizard walks users through setting up differentially private analyses step by step, managing privacy budgets, and generating OpenDP library code. Version 0.5 (2025) added support for linear regression and synthetic data generation. It is explicitly targeted at research data repository users — scientists who need to release aggregate statistics from sensitive datasets for reproducible research without exposing individual records.

Differential Privacy Deployments Registry. Launched in 2025 in collaboration with NIST and the privacy company Oblivious, the registry is a public, collaborative database cataloguing real-world deployments of differential privacy by companies, government agencies, and research institutions. It documents implementation decisions—privacy-loss parameters, privacy units, deployment models, and accuracy trade-offs—to foster shared learning, accountability, and the development of community best practices. The US National Institute of Standards and Technology has proposed hosting the registry; its governance white paper was open for public comment in late 2025. The registry design was informed by a Harvard research study led by postdoctoral researcher Priyanka Nanayakkara, accepted for publication at the IEEE Symposium on Security and Privacy (SP2026).

Tumult Labs acquisition. In 2025, OpenDP took over maintenance and further development of Tumult Labs' open-source DP software, which had been built on a similar programming framework to OpenDP and was recognized for its high bar for correctness and robustness. Former Tumult Labs employees joined the OpenDP team.

Community building. OpenDP runs an annual community meeting at Harvard, an interns and visiting fellows program (now in its fifth year as of 2025), workshops on differential privacy beyond algorithms, and an active Slack workspace and mailing list open to anyone. The 2024 Community Meeting hosted a Differential Privacy Beyond Algorithms Workshop exploring the human, legal, and policy dimensions of DP deployment alongside the technical ones.

Real-World Applications

OpenDP targets use cases in three main sectors. For government statistical agencies, it offers a pathway to release sensitive public data—census figures, income statistics, health records—with formal privacy guarantees replacing weaker pseudonymization or suppression methods. The Swiss Federal Statistical Office has used OpenDP to prototype income statistics releases and developed Lomas, a remote data science platform integrating OpenDP. For academic data repositories such as Dataverse, ICPSR, and Zenodo, it enables researchers to offer privacy-protected access to sensitive datasets that would otherwise be locked away or shared through inadequate de-identification. And for international organizations and NGOs, it provides tools to share humanitarian data responsibly: the UN High Commissioner for Refugees (UNHCR) worked with OpenDP in 2024–25 to generate differentially private synthetic registration datasets for refugee populations across multiple countries—making data accessible to researchers and policymakers while protecting individuals at acute risk. Other documented uses include deployments by Oblivious (for UN and telecom pilot projects) and by OpenMined (for PySyft deployments with Microsoft and DailyMotion, with pilots at national statistical organisations).

Access and Usage

The OpenDP Library is freely available on GitHub and documented at docs.opendp.org. DP Wizard and the Differential Privacy Deployments Registry are accessible at opendp.org. OpenDP’s differential privacy tookit can be found at smartnoise.org. All software is open-source. The community is open to anyone—researchers, data practitioners, government analysts, and engineers — through Slack, GitHub, and the OpenDP mailing list. Those with sensitive data seeking to apply differential privacy can contact OpenDP directly; those with DP research or engineering skills can contribute to the library, DP Wizard, or the deployments registry. Annual fellowship and internship positions are advertised on the OpenDP website each December.