Homomorphic encryption on deep learning in accurate prediction of brain tumour

Abstract

The brain is the most complicated organ that manages every bodily function as well including intellect, memory, emotion, taste, motor skills, vision, respiration, temperature, and appetite. Any type of disease or damage can obstruct the function of the brain and can change the daily lifestyle of a person in an instant or gradually. One of those diseases is a brain tumor, which is hard to detect as serious symptoms start to develop in the later stages of the disease. There are mainly non-cancerous and cancerous brain tumors To make it easier to detect brain tumors we have used the existing Neural Network model to identify tumors. Our objective is to keep patient data confidential as medical institutions are not willing to share patient information due to patients’ rights. And so we have integrated our own Homomorphic encryption so that existing NN models can work and detect tumors from encrypted image datasets. Different Deep Learning and Neural Network techniques can enhance the tumor identification process. In our paper, we have built our custom-made Partial Homomorphic Encryption (PHE) which is based on Paillier Cryptosystem to encrypt the medical data and then used pre-built Neural Networks models (VGG16, VGG19, ResNet50) have been chosen to execute on the data-set consisting of encrypted images of different types of tumors. We have taken the characteristics from the encrypted photos of these brain tumors and extracted them using a pre-trained deep CNN model. First, we have used different Machine Learning algorithms and neural networks in deep learning to classify the images into two categories. Then, we compared the accuracy of various models to identify which algorithm performs the best. For our research, we have created a combined dataset by collecting images of the diseases mentioned above from different sources and applying data augmentation to them. Using our proposed model we can safely and securely read encrypted medical image data via our Partial Homomorphic Encryption method while being efficient enough to be used on a mass scale in the medical industry.