FedHIL: A Novel Indoor Localization Algorithm for Use with Mobile Devices

Preserving Privacy While Enabling Navigation and Emergency Location of People and Devices

At a Glance

Researchers at Colorado State University have developed a novel embedded framework called FedHIL that achieves an average of 1.62 times better indoor localization accuracy than the best performing frameworks previously available. FedHIL preserves user privacy in diverse environments and can be utilized to assist personal navigation, emergency responders, security, and healthcare workers identify the location of individuals or assets within a building.


Indoor localization is the process of determining the location of a device (commonly a mobile device) or user within an indoor environment and is increasingly being used for personal navigation, emergency response, security, and healthcare. Wi-Fi is almost ubiquitously deployed across indoor locations and thus is an integral part of many indoor localization methods. Wi-Fi signals also have the benefit of penetrating through walls and other obstacles, providing a more comprehensive coverage of indoor environments. Machine learning is becoming widely used to improve localization accuracy. However these methods require a large amount of training data and can struggle with different types of mobile devices (device heterogeneity) having variations in Wi-Fi chips and antennas, which can result in variations in signal strength and localization accuracy. There are also concerns about privacy regarding sharing data for training and if the data provided by devices is noisy, it can lead to degradation in model performance. When Wi-Fi routers are replaced, the localization system looses data and accuracy may decrease. More efficient tools are still needed that work with a variety of devices and are not effected by updates to equipment.


The FedHIL framework is a federated learning model, where a central server contains a global model (GM) that is sent to individual mobile devices and trained on the local data, creating a local model (LM). The trained weights for the LM are then sent back to the central server and used to update the GM. In order to help the GM generalize well for different devices, a custom stacked autoencoder is a key feature of the original training that helps improve the accuracy of the model. It is combined with a shallow neural network, that is a lightweight model ideal for mobile devices. When aggregating the LM and GM, only a subset of the LM weights (high impact weights) are selected to protect the GM from noisy, low impact data and preserve the accuracy of the model over multiple rounds of retraining. Additionally, FedHIL preserves user privacy by only updating a subset of the high impact weights from the LM to the GM, preventing the need to update location details of the mobile devices and ensuring the privacy of the mobile device. When the FedHIL framework was compared against other federated and non-federated learning models of indoor localization, FedHIL had the lowest average error of 3.24 meters and performed 1.62 times better than the best federated learning model.


  • At least 1.62 times better average accuracy over other methods
  • Average error of 3.24 meters
  • Preserves user privacy
  • Works in diverse environment and with diverse mobile devices
  • Robust model resilient to noise


  • Emergency response
  • Security
  • Healthcare
  • Indoor navigation
Last Updated: February 2024
phone showing a map in a store

Available for Exclusive Licensing
TRL: 4


Sudeep Pasricha
Danish Gufran

Reference Number
Licensing Manager

Aly Hoeher