Differential privacy medical image classification with deep learning method

Nowadays, with the growing use of deep learning applications, the size and variety of data used in these applications have increased significantly. For this reason, in addition to the studies focusing on model performance, various studies also have been started to provide the security of deep learning models. In studies that provide the privacy of model training data, differential privacy applications come forward. Because differential privacy can be applied to every problem and, the tradeoffs between throughput and privacy is accounted. In this thesis, a differentially private deep learning model was developed on medical images which contain sensitive personal data such as the health status of individuals. After that, the performance of the model was tested. In this study, multiscale 3D convolutional neural network is used as a model with DP-SGD optimization algorithm based on gradient perturbation. In the developed model, diagnosis of patients with brain tumor or cancer was made with LGG and HGG glioma classification in the tumor grading step, at the same time respecting privacy. The BraTS 2020 dataset, which includes brain MR images, was used as the training and test set. The performance rate of privacy-sensitive and insensitive deep learning models was compared, and it was seen that the model performances did not change to a great extent. With an eye on that the model performance is around 85% on average, it is considered that higher accuracy will be achieved when the dataset size increases. With this, the brain MR images can be analyzed with respecting privacy and differential privacy can be applied to the datasets within the scope of the Turkish Brain Project.