METHOD AND APPARATUS FOR TRAINING GBDT MODEL IN TRUSTED EXECUTION ENVIRONMENT, AND DEVICE

WIPO Patent Application WO/2021/098385

A1

A method and apparatus for training a GBDT model in a trusted execution environment, and a device. In the method, each feature in a D-dimensional feature is used as the current feature in sequence, and node division determining is performed on each node in the current level in the current decision tree, the node segmentation determining comprising: loading N encrypted arrays of the current feature to a trusted memory area and decrypting the N encrypted arrays; using each of obtained N original arrays as the current array in sequence, and determining a corresponding first node; and on the basis of the current feature value in the current array, performing hypothetical segmentation on the sample set corresponding to the first node, and on the basis of the hypothetical segmentation result, determining the current best split feature and current best feature threshold of the first node; after the node segmentation determining is performed on the basis of each feature in the D-dimensional feature, the sample set of each node is segmented to generate the next level of nodes until a leaf node is reached. Thus, security training of a GBDT model can be implemented on the basis of private data.

ZHAO YUAN (CN)
YIN SHAN (CN)

PCT/CN2020/118507

May 27, 2021

September 28, 2020

Click for automatic bibliography generation  

ALIPAY HANGZHOU INF TECH CO LTD (CN)

G06N20/10

CN111126628A	2020-05-08
CN107301577A	2017-10-27
CN109308418A	2019-02-05
CN108694673A	2018-10-23
US20190318202A1	2019-10-17

BEIJING BESTIPR INTELLECTUAL PROPERTY LAW CORPORATION (CN)

View/Download PDF      PDF Help

Previous Patent: DATA ABNORMALITY DETECTION METHOD AND APPARATUS

Next Patent: MULTIFUNCTIONAL RAIL TRANSPORTATION AND LAYING MACHINE FOR MAGLEV TRAIN