US 12,135,784 B2
Privacy-protection-based data processing model acquisition method and apparatus, terminal device and storage medium
Xin Huang, Beijing (CN)
Assigned to ENNEW DIGITAL TECHNOLOGY CO., LTD, Beijing (CN)
Filed by ENNEW DIGITAL TECHNOLOGY CO., LTD, Beijing (CN)
Filed on Apr. 1, 2022, as Appl. No. 17/711,048.
Application 17/711,048 is a continuation of application No. PCT/CN2021/101138, filed on Jun. 21, 2021.
Claims priority of application No. 202010778938.1 (CN), filed on Aug. 5, 2020.
Prior Publication US 2022/0229904 A1, Jul. 21, 2022
Int. Cl. G06F 21/55 (2013.01); G06F 21/60 (2013.01); G16Y 40/50 (2020.01)
CPC G06F 21/556 (2013.01) [G06F 21/60 (2013.01); G16Y 40/50 (2020.01); G06F 2221/034 (2013.01)] 8 Claims
OG exemplary drawing
 
1. A privacy-protection-based data processing model acquisition method, comprising:
acquiring sensor data of a plurality of sensors of a preset Internet of Things device;
training an initial data model corresponding to each of the sensors through the sensor data corresponding to the sensor to obtain an intermediate data model corresponding to each of the sensors, and integrating the intermediate data models corresponding to the sensors to form an integrated data model;
processing new data through the integrated data model and random noise to acquire a label category corresponding to the new data; and
training the integrated data model according to the new data and the label category of the new data to acquire a data model,
wherein, the step of processing new data through the integrated data model and random noise to acquire a label category corresponding to the new data comprises:
inputting the new data to each of the intermediate data models in the integrated data model to acquire a label category outputted by each of the intermediate data models;
determining a number of initial notes of each of the label categories according to the label category;
introducing random noise on the basis of the number of initial notes to acquire a number of notes of each of the label categories; and
determining, according to the number of notes of each of the label categories, the label category whose number of notes meets a preset condition as the label category of the new data; and
wherein, the random noise comprises Laplacian noise, the Laplacian noise is Lap(1/ε), ε denotes privacy costs, and a density function thereof is

OG Complex Work Unit Math
where x denotes the new data, and b=1/ε.