Research Article
Privacy-Preserving Decision Tree Classification Protocol Based on Bitwise Comparison
@INPROCEEDINGS{10.1007/978-3-031-23902-1_8, author={Peihang Yu and Baodong Qin and Dong Zheng}, title={Privacy-Preserving Decision Tree Classification Protocol Based on Bitwise Comparison}, proceedings={Mobile Multimedia Communications. 15th EAI International Conference, MobiMedia 2022, Virtual Event, July 22-24, 2022, Proceedings}, proceedings_a={MOBIMEDIA}, year={2023}, month={2}, keywords={Machine learning Decision tree Secret sharing Homomorphic encryption Privacy-preserving}, doi={10.1007/978-3-031-23902-1_8} }- Peihang Yu
Baodong Qin
Dong Zheng
Year: 2023
Privacy-Preserving Decision Tree Classification Protocol Based on Bitwise Comparison
MOBIMEDIA
Springer
DOI: 10.1007/978-3-031-23902-1_8
Abstract
Decision tree model is widely used in telemedicine, credit evaluation and other fields because of its high efficiency and ease of use. Companies can charge customers who do not have professional knowledge or resources to build pre-diction models to achieve the purpose of profitability. In order to protect the privacy of client data and decision tree model parameters in this scenario, we propose an integer comparison protocol based on distributed bitwise comparison, and further design an efficient privacy-preserving decision tree classification model. Our protocol uses the idea of Beaver triple multiplication combined with secret sharing to replace the ciphertext homomorphic operation in the original comparison protocol. Compared with the comparison protocol re-lying on the traditional semi-homomorphic encryption scheme, it further im-proves the operation efficiency and compresses the communication cost in ciphertext transmission. Security analysis shows that this scheme achieves the expected privacy for users and model providers. Experimental results on real datasets show that the overall running time of this scheme is about 0.8s for a decision tree of depth 3, and about 5s for a decision tree of depth 17. So, its computational overhead is low and acceptable.
