Federated Learning for Localization: A Privacy-Preserving Crowdsourcing Method
share
Received Signal Strength (RSS) fingerprint-based localization attracted a lot of research effort as well as cultivating many commercial applications of location-based services due to its low cost and ease of implementation. There are many studies exploring use of machine learning algorithms for RSS fingerprint-based localization. However, there are several problems due to the brittle nature of RSS which is susceptible to noise, dynamic environments, and hardware variations. These problems require frequent retraining of machine learning models with new data gathered by crowdsourcing. Although crowdsourcing is a good way to collect data, it has a privacy problem due to collection of raw data from participants. Federated learning enables to preserving privacy of the crowdsourcing participants by performing model training at the edge devices in a decentralized manner, data of participants is not exposed to centralized entity. This paper presents a novel method to improve the reliability and the robustness of RSS fingerprint-based localization while preserving privacy of the participants and the users by employing federated learning. Proposed method could achieve below 1 meters difference using federated learning compared to centralized learning for RSS fingerprint-based localization, and improved localization accuracy by 16
READ FULL TEXT