CSITime: Privacy-preserving human activity recognition using WiFi channel state information

Area of Science:

• Computer Science
• Artificial Intelligence
• Signal Processing

Background:

• Human Activity Recognition (HAR) is crucial for smart applications but faces limitations with traditional vision and sensor methods.
• WiFi Channel State Information (CSI) offers a promising, cost-effective alternative for HAR due to easy deployment.
• Existing CSI-based HAR methods require further optimization for accuracy and robustness.

Purpose of the Study:

• To propose CSITime, a modified InceptionTime network for enhanced CSI-based human activity recognition.
• To treat CSI activity recognition as a multi-variate time series problem.
• To improve the accuracy and robustness of WiFi-based HAR systems.

Main Methods:

• Developed CSITime, a network integrating convolutional kernels, self-attention, and Mish activation.
• Employed data augmentation techniques like mixup and cutmix for improved neural network learning.
• Utilized one-cycle policy and cosine annealing for efficient neural network training.

Main Results:

• CSITime achieved state-of-the-art accuracies: 98.20% (ARIL), 98% (StanWiFi), and 95.42% (SignFi).
• Demonstrated superior performance over existing methods by 3.3%, 0.67%, and 0.82% on benchmark datasets.
• Showcased improved robustness, achieving 2.17% higher accuracy on a challenging SignFi dataset scenario with distribution shifts.

Conclusions:

• CSITime represents a significant advancement in CSI-based human activity recognition.
• The proposed architecture and training strategies effectively enhance HAR performance.
• CSITime offers a robust and accurate solution for real-world HAR applications.