Jin, C., wang, J., Teo, SG., Zhang, L., Chan, CS., Hou, Q., Aung, KMM. Towards End-to-End Secure and Efficient Federated Learning for XGBoost. fl-aaai-2022
Abstract:
Federated learning refers to the distributed and privacypreserving collaborative machine learning paradigm, in
which multiple independent data owners jointly train on certain machine learning models without revealing their private
data information to each other. In this paper, we study federated learning on XGBoost models in vertical data partition
settings where the data owners share a common set of training samples, and each data owner possesses a disjoint subset
of the features. We propose CryptoBoost, a federated XGBoost system based on multi-party homomorphic encryption
techniques. CryptoBoost outperforms previous works in majorly three aspects. 1) CryptoBoost is end-to-end secure that
the models are trained and stored in a completely encrypted
and private manner. 2) CryptoBoost eliminates any central or
privileged node that knows or controls more information than
the other nodes, and the federated learning and inference processes are done in a fully decentralized way. 3) We propose
a set of new secure computation algorithms and protocols for
CryptoBoost, which achieve improved performance and communication efficiency compared with existing approaches.
License type:
Publisher Copyright
Funding Info:
This research / project is supported by the RIE2020 - Advanced Manufacturing and Engineering (AME) Programmatic Programme
Grant Reference no. : A19E3b0099