search

mc2-project / secure-xgboost

Secure collaborative training and inference for XGBoost.
https://mc2-project.github.io/secure-xgboost/
Apache License 2.0
105 stars 32 forks source link

Can this work be federated？ #169

Open njuptlht opened 1 year ago

njuptlht commented 1 year ago

I read your another work "federated xgboost", which stated that it does not support differential privacy. If I implement this work in a distributed system, can it be considered federated？

chester-leung commented 1 year ago

Hi @njuptlht thanks for your interest in our project.

The Secure XGBoost model is that of multiparty outsourced computation, i.e. where all parties' data are individually encrypted and entirely transferred to a central cluster for processing. This is different from federated learning, in which only data summaries are exchanged between parties and a central aggregator service, and in which (in its vanilla form) summaries are unencrypted. The XGBoost training algorithm, however, natively exchange summaries -- see section 3 of the paper here for more.

If you are looking for differential privacy, this project also does not yet support it.