| Autor: |
Xu, Yang, Hu, Xuexian, Wei, Jianghong, Yang, Hongjian, Li, Kejia |
| Předmět: |
|
| Zdroj: |
Journal of King Saud University - Computer & Information Sciences; Jan2023, Vol. 35 Issue 1, p237-249, 13p |
| Abstrakt: |
With growing concerns about privacy and the fact that data are distributed among multiple parties in realistic scenarios, vertical federated learning (VFL) is becoming increasingly important. There is an increasing trend in adapting machine learning algorithms to the VFL setting. As a category of prevalent machine learning algorithms, decision tree and random forests in VFL have attracted widespread interest. However, existing frameworks suffer either from potential privacy breaches or high communication consumption. To close this gap, we propose a communication-efficient vertical federated framework for the classification and regression tree (CART) algorithm called VF-CART, and extend it to random forests (RFs). Specifically, we convert feature values into bin values and build a histogram for each feature. By employing a hash function and homomorphic encryption to secretly choose the best split, a participant with labels cannot obtain the sample subsets for each split. In addition, the number of ciphertexts transmitted between entities is reduced significantly in both the training and prediction stages. Participants who do not have the labels communicated only once with the third-party server during the tree-building stage. During the prediction stage, only one ciphertext must be transmitted to predict a sample. Finally, we conducted experiments using both real-world and synthetic datasets. The experimental results demonstrate that the VF-CART algorithm significantly reduced the volume of communication. [ABSTRACT FROM AUTHOR] |
| Databáze: |
Supplemental Index |
| Externí odkaz: |
|
