INTERNATIONAL JOURNAL OF ADVANCE RESEARCH
IN MULTIDISCIPLINARY

Machine learning algorithms may be trained on decentralized data using Federated Learning (FL) when sharing raw data is not possible owing to privacy concerns. One example of this kind of data is EHRs, or electronic health records, which store private information about patients. Instead of sharing sensitive data, FL trains models locally and then aggregate their parameters on a central server. An effective method for training Machine Learning (ML) algorithms on distributed datasets when data owners are governed by restrictions that limit the sharing of raw data is Federated Learning (FL). There is less need to communicate raw data with people outside the premises with this strategy, which involves local training and model aggregation to a central server. Nevertheless, FL brings up valid issues around privacy. For that reason, we need more privacy safeguards. One state-of-the-art privacy technique is the differential privacy (DP) approach, which involves adding an extra layer of privacy by perturbing the local models before transmission. But this method could change the framework's usefulness. In order to strike a fair balance between privacy and usefulness, we employ a private method to clean raw data by combining DP noise with a top-down taxonomy tree. To train local models that may be shared in the FL architecture, the generalized data is utilized in conjunction with DP noise. The suggested architecture improves functionality while keeping the privacy budget low.