Proposal: adding a section in the documentation for privacy-preserving XGBoost

[bcm-at-zama]

Hello XGBoost community, and congratulations for what you are building, it's awesome.

We at Zama are working on making a privacy-preserving ML open source library, providing the same kind of APIs than sklearn, torch, or... XGBoost, with the important difference that data are encrypted through the whole inference process, i.e., never decrypted. It's possible, thanks to so-called Fully Homomorphic Encryption. Our library is called Concrete ML.

Of course, we have an XGB implementation working over encrypted data, see eg this example. We have also deployed it in an Hugging Face Space.

The reason to make this issue was to ask you if we could add a link to our project in your documentation. Maybe over a new "Privacy Preserving Prediction", below the current Prediction, or some other place you would tell me? The goal is to increase the visibility of the library, and encourage more users to go for privacy-preserving solutions, which is a must for some sensible data.

If you're good with it, I can prepare some PR. Thank you.

[trivialfis]

cc @rongou

[rongou]

I think so far the documentation is mostly about features in XGBoost itself. There are tons of external integration points that we can't possibly cover. That being said, I'm ok with adding a pointer to your work, something along the line of "for privacy preserving prediction, see Concrete ML". Probably shouldn't add a complete example, as it could quickly go out of date.

We are also looking at FHE for federated learning, especially with GPU acceleration (see blog post).

[bcm-at-zama]

Great, thanks a lot. So, should I make a PR (you would tell me where to add the few new lines) or I let you take care of everything? I think the links https://github.com/zama-ai/concrete-ml/blob/main/docs/advanced_examples/XGBClassifier.ipynb and https://github.com/zama-ai/concrete-ml/blob/main/docs/advanced_examples/XGBRegressor.ipynb will be permanent, or you have https://docs.zama.ai/concrete-ml/built-in-models/tree.

Anyway: thanks a lot for consideration and quick answer, I'm always amazed by open-source community

[rongou]

A PR would work. Do you have some kind of landing page to provide a high level overview?

[bcm-at-zama]

Sure, let me prepare a PR. Which file to you want me to modify / where should I add the links, please?

For the landing page, I was thinking about:

• our GitHub: https://github.com/zama-ai/concrete-ml
• our documentation: https://docs.zama.ai/concrete-ml
  but if you want to precisely link to somewhere speaking about XGB, the previously cited examples would make sense as well.

[rongou]

Probably this file: https://github.com/dmlc/xgboost/blob/master/doc/prediction.rst

The docs page should be good.

[trivialfis]

Let's add the document to the doc/tutorials/ instead. ;-)

[bcm-at-zama]

Hello @rongou. Yes, done in #9604, thanks.

[bcm-at-zama]

Regarding your other proposal @trivialfis, if you propose us to have a Concrete ML example in doc/tutorials/ and @rongou is fine with it, of course, we will! Tell us, thank you

[trivialfis]

I thought you wanted to add a full tutorial for using it. But if that's a brief introduction with links to your repository for more info, then the current PR is fine.

[bcm-at-zama]

@trivialfis : actually, it's also an excellent suggestion, and my colleague https://github.com/jfrery told me he volunteered to do such a tutorial. It will come, in another PR than #9604

[trivialfis]

Hi, both PRs are now merged, thank you for the nice work! Is there anything else XGBoost needs to do?

[bcm-at-zama]

Hello @trivialfis . Yes, we'll close this issue with #9614. We just missed to add the new tutorial in the index. Cheers