Research On Privacy-preserving Classification Methods For Medical Image

The rapid development of medical digital equipment produced a large number of medical images.Using deep learning methods to classify medical images became a research hotspot.However,in this process,there is a serious problem of privacy disclosure.This is because the third-party responsible for medical image management may not be trusted.Medical images contain a large number of patients’ private information.Once leaked,accurate recommendations,advertising fraud and other phenomena will be common,which will cause trouble to patients.There are two common modes of medical image processing.One is to classify medical images on the local centralized database,and the other is to classify medical images from multi-site databases.This thesis starts from the above two models and studies their privacy protection methods in them.Consider using a homomorphic encryption algorithm to encrypt medical images on the local centralized database to ensure image security.In addition,the privacy leakage of the image classification model in the classification process is also considered,and the model is also homomorphically encrypted.The difficulty lies in classifying and ensuring classification accuracy when both medical images and image classification models are encrypted.For medical images from multi-site databases,consider using a federated learning model to allow local sites to independently complete model training,and then upload parameters to a central server for secure aggregation.The urgent problem to be solved is to prevent malicious users from obtaining training data through model inversion and to ensure the generalization ability of the model.This thesis proposes a medical image classification method with privacy protection given the above problems.The main work is as follows:Firstly,aiming at the privacy leakage problem that may be caused in the process of medical image classification on the local centralized database,this thesis proposes a secure classification scheme of encrypted medical images based on a convolutional neural network.The scheme uses Paillier encryption to encrypt the convolutional layers,pooling layers,activation layers,and fully connected layers of medical images and convolutional neural networks,and designs related privacy protection functions.The use of privacy-preserving functions enables image classification where both medical images and image classification models are ciphertexts.Experiments were performed on real chest X-rays,malaria cells,and blood cell image sets,and achieved over 94%accuracy.Analysis shows that this scheme can maintain the same accuracy as the plaintext model.Secondly,aiming at the privacy leakage problem that may be caused in the process of medical image classification from multi-site databases,this thesis proposes a differential privacy medical image classification method based on federated learning.The method combines a differential privacy mechanism with federated learning to ensure the privacy and security of medical images,and the addition of random noise can prevent malicious users from inferring the model to obtain the original image.During the training process of the client’s local model,a new adaptive gradient descent algorithm is introduced to ensure the convergence of the model and improve the generalization ability of the model.Experimental results on chest X-ray images show that this method has higher accuracy,faster convergence speed and better generalization ability,and ensures the safety of medical images.