Federated learning empowers preparing a global accepted machine learning model from information disseminated over various locations, without moving the information. This is very crucial in health related applications, where information is overflowing with individual, exceptionally sensitive data, and diagnostic tools should provably adhere to governmental rules. Although federated learning envisions sharing raw information, it is as yet conceivable to report privacy assaults on the model parameters that are uncovered during the preparation procedure, or on the created machine learning model. In this paper, we propose the (k,l)f-anonymity model for offering privacy with regards to federated learning approach. Unlike the differential privacy-based systems, our modeltries to boost utility or model execution, while supporting a greater degree of privacy, as requested by GDPR and HIPAA. We run through a thorough observational assessment on significant issues in the healthcare area, utilizing genuine electronic wellbeing information of one million patients. The outcomes show the adequacy of our methodology in accomplishing high model execution, while offering the ideal degree of privacy. Through relative investigations, we likewise show that, for changing datasets, trial arrangements, and privacy budgets, our methodology in federated learning offers higher model execution than the centralized setting.
Volume 12 | 06-Special Issue
Pages: 250-262
DOI: 10.5373/JARDCS/V12SP6/SP20201030