Explain why you should choose OpenDP

[mccalluc]

There are multiple DP libraries out there. Near the top of the docs we should have a short argument for why someone should pick OpenDP, perhaps bullet points or a feature matrix. (And if other libraries do something well, acknowledge that, too! Having a presentation that is balanced can help people feel more confident that they've done their research.)

Some possible points:

• Under active development.
• Multiple language bindings give you the same functionality in the language of your choice.
• Rigorous development process (Rust backend, algorithm reviews) has produced a codebase that performs well in independent security audits (ie Group and Attack: Auditing Differential Privacy)
• Broad API supports a range of DP mechanisms
• (when Context API is out) Relatively easy to use

On the other side:

• OpenDP is agnostic about where the data is coming from. For any particular backend, there may be a library that provides some elements of DP with minimal changes to existing code.
  • https://pipelinedp.io/ -- "Provides a convenient API familiar to Spark or Beam developers"
  • TensorFlow Privacy
  • Opacus for PyTorch
  • SQL integrations

Should also compare with:

• https://github.com/IBM/differential-privacy-library
• https://github.com/OpenMined/PyDP
  • Built on https://github.com/google/differential-privacy

Please add to the list!

[SalilVadhan]

Backed by and (eventually) governed by the community (as opposed to a single company)

---

Salil Vadhan (he/they)

On Wed, Nov 15, 2023 at 6:23 AM Chuck McCallum @.***> wrote:

There are multiple DP libraries out there. Near the top of the docs we should have a short argument for why someone should pick OpenDP, perhaps bullet points or a feature matrix. (And if other libraries do something well, acknowledge that, too! Having a presentation that is balanced can help people feel more confident that they've done their research.)

Some possible points:

• Under active development.
• Multiple language bindings give you the same functionality in the language of your choice.
• Rigorous development process (Rust backend, algorithm reviews) has produced a codebase that performs well in independent security audits (ie Group and Attack: Auditing Differential Privacy https://www.sri.inf.ethz.ch/publications/lokna2023groupandattack)
• Broad API supports a range of DP mechanisms
• (when Context API is out) Relatively easy to use

Please add to the list!

— Reply to this email directly, view it on GitHub https://github.com/opendp/opendp/issues/998, or unsubscribe https://github.com/notifications/unsubscribe-auth/AB4KIGO63YNS6HTO5CR26UDYETF6HAVCNFSM6AAAAAA7MSL366VHI2DSMVQWIX3LMV43ASLTON2WKOZRHE4TIOBXHE3DOMA . You are receiving this because you are subscribed to this thread.Message ID: @.***>

[mccalluc]

Before making the case for OpenDP, we should make the case for DP in general:

• See #317

[mccalluc]

Work has been done in #1065, but that's going to live in a feature branch for a while before we merge to main and close this issue.

[mccalluc]

The new docs on main cover this:

Why OpenDP?

• OpenDP is based on a solid conceptual framework for expressing privacy-aware computations.
• OpenDP is built on a Rust core for memory and thread safety and performance.
• OpenDP has a process for independent review of algorithms and implementations.
• OpenDP has performed well in independent security audits.
• OpenDP supports a range of differential privacy algorithms.
• OpenDP has bindings for Python and R, both built on the same Rust core for consistency and security.

• OpenDP is a community effort and is not owned or directed by a single corporation.

  That said, OpenDP is not the best tool for every job. In particular, it is a fairly low-level interface: There are a number of other projects which try to make it easy to add differential privacy to existing SQL interfaces or ML frameworks. One such tool is SmartNoise SDK, which is built on the OpenDP library.

Don't need to wait for release to close.