Institution profile

Latitude AI

Industry researchnorthamerica · us
Official website
Research library1linked papers
Opportunities0open roles
Selected work

Representative Papers

FedWiLoc: Federated Learning for Privacy-Preserving WiFi Indoor Localization

Dec 19, 2025

To address three key challenges in Wi-Fi indoor localization—privacy leakage, degraded accuracy under multipath conditions, and poor cross-scenario generalizability—this paper proposes a privacy-first federated CSI-based localization framework. Our method employs access point (AP)-side local CSI encoding and embedding vector uploading, integrated with split neural networks and federated learning for collaborative modeling. We further introduce an angle-position joint geometric consistency loss to enforce physical constraints during training. Evaluated across six real-world indoor environments (>2000 sq. ft.), our approach reduces median localization error by 61.9% compared to state-of-the-art methods. Crucially, it guarantees end-to-end privacy preservation throughout both training and inference phases—without requiring raw CSI data sharing. To the best of our knowledge, this is the first work to simultaneously achieve high accuracy, strong privacy guarantees, and broad generalizability across diverse deployment scenarios.

0 citationsRead paper
Recent publications

Latest Papers

FedWiLoc: Federated Learning for Privacy-Preserving WiFi Indoor Localization

Dec 19, 2025

To address three key challenges in Wi-Fi indoor localization—privacy leakage, degraded accuracy under multipath conditions, and poor cross-scenario generalizability—this paper proposes a privacy-first federated CSI-based localization framework. Our method employs access point (AP)-side local CSI encoding and embedding vector uploading, integrated with split neural networks and federated learning for collaborative modeling. We further introduce an angle-position joint geometric consistency loss to enforce physical constraints during training. Evaluated across six real-world indoor environments (>2000 sq. ft.), our approach reduces median localization error by 61.9% compared to state-of-the-art methods. Crucially, it guarantees end-to-end privacy preservation throughout both training and inference phases—without requiring raw CSI data sharing. To the best of our knowledge, this is the first work to simultaneously achieve high accuracy, strong privacy guarantees, and broad generalizability across diverse deployment scenarios.

0 citationsRead paper