Privacy Preservation of Medical Images Through Synthesized Federated Learning Framework in Healthcare

The correctness of AI depends on the correctness of the data.Although AI has brought numerous developments in the healthcare industry, we can't deny the fact that there are security breaches as well.This synthesized framework is proposed as a solution for providing security and privacy.Medical images of Alzheimer's disease are considered in this study.Medical images occupy huge space and also consume more bandwidth while being transmitted.A distributed learning model called federated learning is employed that allows the images to reside at any hospital.Training is performed at the client end itself and only the model parameters are shared to the server.It preserves privacy using Partially Homomorphic Encryption and the encrypted images are used for training.This proposed model uses FedAvg for model aggregation.Two CNN architectures are considered here-ResNet50 and DenseNet121.The experimental results show that the DenseNet121 model gives more accuracy than ResNet50 for the encrypted image dataset.