Hyper-Noise Interference Privacy Protection Framework for Intelligent Medical Data-Centric Networks

Abstract

Convolutional neural network (CNN) has been widely used in the medical domain, but how to protect privacy is still a challenge. In addition, traditional differential privacy protection methods are very sensitive to noise; hyper-scale noise has a serious effect on the validity of the network. In this article, therefore, a novel framework named improved private aggregation of teacher ensembles (IPATE) is proposed. IPATE aims to protect the privacy of trained models and medical data by adding hyperscale Laplace noise to the extended voting difference of many CNN models. An experimental example including 12 leads data of 148 myocardial infarction patients and 52 normal subjects was used in the study. Experimental results show that with the increase of noise and under the condition of limited data, IPATE can still achieve about 90 percent accuracy, and it is not sensitive to noise.

Ähnliche Arbeiten