Critical Analysis of Data Leakage in WiFi CSI-Based Human Action Recognition Using CNNs

Area of Science:

• Computer Science
• Electrical Engineering
• Signal Processing

Background:

• WiFi Channel State Information (CSI) is used for non-intrusive human action recognition.
• Convolutional Neural Networks (CNNs) are a common method for this task.
• Concerns exist regarding the reliability of reported performance metrics due to potential data leakage.

Purpose of the Study:

• To critically analyze a specific IEEE Sensors Journal study on WiFi CSI-based human action recognition.
• To identify and demonstrate data leakage issues in the original study.
• To reassess performance metrics using correct, subject-based data partitioning.

Main Methods:

• Analysis of data partitioning methods in WiFi CSI human action recognition.
• Empirical investigation to verify the lack of individual exclusivity across dataset partitions.
• Re-evaluation of performance metrics with subject-based data splitting.

Main Results:

• Instances of data leakage were found in the analyzed study due to improper data partitioning.
• Empirical evidence confirmed that individuals were present in multiple dataset partitions.
• Subject-based partitioning led to significantly lower precision rates than previously reported, refuting the 99.9% claim.

Conclusions:

• The original study's high performance metrics were inflated due to data leakage.
• Rigorous data management, specifically subject-based partitioning, is crucial for reliable WiFi CSI-based activity monitoring.
• Accurate reporting is essential to prevent misleading results and foster genuine research progress.