ACDP-Floc: An Adaptive Clipping Differential Privacy Federation Learning Method for Wireless Indoor Localization

With the increasing demand for location services, the fingerprint recognition technology based on the received signal strength (RSS) has been paid more and more attention and applied due to its advantages of mature infrastructure and easy implementation, Federated Learning (FL) has been applied to indoor localization to solve data silos and privacy security problems in recent research work. To prevent eavesdroppers from inferring private information and model features of the client by analyzing parameter information. Some researchers introduce differential privacy (DP) technology into FL for privacy protection, but the addition of noise seriously affects the availability of data and models. We investigate the privacy loss measurement and tracking methods of DP and propose ACDP-Floc, an adaptive clipping differential private federated learning method for indoor location, the usability of data and model is improved by adaptive clipping of model gradient. Experimental results show that: when the privacy budget \(\varepsilon =1.0\) , which indicates that the algorithm adds a large noise, ACDP-Floc achieves 92.53%, 93.61% and 96.54% classification accuracy for the Mall Area, Mall-Wi-Fi and UIJIIndoorLoc three real datasets, respectively.