Fast-Adapting Environment-Agnostic Device-Free Indoor Localization via Federated Meta-Learning

Deep learning-based device-free fingerprinting indoor localization faces the challenge of high data-labeling and training costs, especially when localization is required in multiple environments. A general model that can adapt to multiple environments and reduce these costs while maintaining data privacy is highly desirable. This paper proposes a federated metalearning framework for device-free indoor localization, where each client, representing an environment or task, collaboratively train a general environment-agnostic model while preserving their data privacy. Fast adaptation to new environments is achieved by downloading the general model from the server and updating the model locally with only few labeled data. The proposed system is applicable to heterogeneous environments with varying layouts, dimensions, or numbers of locations. Real-world experiments demonstrate the effectiveness of the proposed method and its potential for significant data-labeling and training cost reductions.