SenseFi: A library and benchmark on deep-learning-empowered WiFi human sensing.

Area of Science:

• Computer Science
• Signal Processing
• Artificial Intelligence

Background:

• WiFi sensing is advancing for privacy-preserving human-sensing using signal processing and deep learning.
• A public benchmark for deep learning in WiFi sensing is currently lacking, hindering research progress.

Purpose of the Study:

• To introduce SenseFi, a comprehensive benchmark and open-source library for deep learning in WiFi sensing.
• To evaluate deep learning models on diverse sensing tasks, WiFi platforms, and performance metrics.

Main Methods:

• Review of recent advancements in WiFi sensing hardware and algorithms.
• Development and implementation of the SenseFi benchmark library.
• Extensive experimental evaluation of various deep learning models within SenseFi.

Main Results:

• SenseFi provides a standardized platform for comparing deep learning models in WiFi sensing.
• Experiments offer insights into model design, learning strategies, and training for real-world applications.
• Evaluations cover recognition accuracy, model size, computational complexity, and feature transferability.

Conclusions:

• SenseFi serves as a valuable, open-source tool for the WiFi sensing research community.
• The benchmark facilitates the validation of learning-based WiFi-sensing methods across multiple datasets and platforms.
• This work aids in advancing the development of robust and efficient deep learning models for WiFi sensing.