Deep Learning With Differential Privacy And Applied To Medical Image Analysis

The success of the deep learning model largely depends on the quality and quantity of training data.This is particularly problematic in medical image analysis applications,where high-quality data is sparse and data usage is limited due to patient privacy concerns.As a powerful privacy protection technology,differential privacy can provide privacy protection for patients,and can be implemented in deep neural network during training by using differentially private stochastic gradient descent(DPSGD)algorithm.This dissertation focuses on the medical image analysis based on differentially private deep learning.In order to improve the quality of the private model and the stability of training,an improved adaptive differentially private optimization algorithm is proposed.The concrete work describes as follows.Firstly,a faster adaptive algorithm Nadax is proposed by modifying the classical momentum term of the first moment of the Ada X algorithm.On the one hand,it can overcome the shortcoming of SGD algorithm which is sensitive to learning rate parameter.On the other hand,it has comparable generalization ability as SGD algorithm,and can achieve faster convergence speed than other adaptive algorithms.In differentially private deep learning,it means that under the same accuracy requirements,model training requires fewer iterations,thus saving the privacy budget.Then,the adaptive differentially private optimization algorithm S-DPNadax is proposed by using gradient perturbation technique to deploy differential privacy mechanism on Nadax algorithm.On the one hand,aiming at the training stability and privacy-utility balance of the private model,randomized smoothing technique is introduced to obtain the optimization target with better noise tolerance.On the other hand,the advanced Gaussian differential privacy framework is used to track the privacy budget.Compared with other differential privacy frameworks,it can track the privacy budget consumption more finely.In differentially private deep learning,it means that under the same privacy budget,the model can have more iterations,thus achieving higher model prediction accuracy.Finally,the performance of S-DPNadax algorithm in medical image classification and semantic segmentation tasks is evaluated on real datasets.The results show that it is possible to train neural networks with strict privacy protection while maintaining acceptable classification and segmentation performance.In addition,under the same privacy budget,the model accuracy achieved by S-DPNadax algorithm is better than other related work.Figure [14] Table [8] Reference [76]...