Research On The Prediction Of Lung Cancer Disease Based On Privacy Protection Measures Of Federated Learning

Globally,lung cancer is the most commonly diagnosed cancer.The early stage of lung cancer is often presented in the form of lung nodules,so in order to effectively predict lung cancer disease,studies are usually carried out based on the lung nodule detection.With the rapid development of artificial intelligence,neural network has been widely used in the field of lung nodule detection.At the same time,user privacy and data security have been widely concerned.Institutions such as hospitals often have limited access to user data due to the particularity of the industry.In the face of the need to protect user privacy and data security,it is difficult for different medical institutions to aggregate all data together,which gradually forms data islands.To solve the above problems,this dissertation carried out a study on the prediction of lung cancer disease based on the privacy protection countermeasures of federated learning,which has important theoretical significance and broad application prospect for the prevention of lung cancer disease.The main contents are as follows:1.In the increasingly strict legal environment,this dissertation proposes a 3D Res Net18 dual path Faster R-CNN lung nodule detection based on federated learning algorithm to solve the problem of small and fragmented data islands in medical data sets.This algorithm does not need to aggregate the data sets between different medical institutions,but jointly constructs the machine learning model through the federated learning framework,which provides a new idea for lung nodule detection without large data sets for lung nodule.The experimental results show that the 3D Res Net18 dual path Faster R-CNN based on federated learning algorithm can significantly improve the detection performance of lung nodules.2.This dissertation studies the influence of data quality on the model and proposes a data diversity algorithm based on sampling.This algorithm uses a hash function to map local data into a separate bucket,and extracts the content differences among the calculated data in a partial sample set.This algorithm maintains the content diversity and reduces the training time of the model.3.In this dissertation,a dual mechanism differential privacy federated learning algorithm is proposed to solve the problem that semi-honest adversaries can steal data privacy by participating in the federated learning training process and malicious adversaries can steal data privacy by directly attacking clients.This algorithm uses Laplacian mechanism and Gauss mechanism for mixed difference and retains some original data information,which effectively solves the problems of weak privacy protection of federated learning data and the difference privacy algorithm reduces the detection performance of neural network.This algorithm has a potential application prospect.