Description
Anomaly detection flags samples that deviate from a learned model of "normal" data — via reconstruction error (autoencoder, GAN), density (one-class SVM, isolation forest), or distance to a learned manifold. In WiFi sensing it covers fall detection in elderly-care contexts, intrusion detection in smart homes, and drift detection between BLE-CSI calibration campaigns. It is a lightweight alternative to fully labelled multi-class classification when the rare event is the whole point.
When it's used
- Drift detection between calibration epochs
- Intrusion / occupancy-when-empty alerts
- Open-set rejection for HAR / gesture pipelines
- Sensor-health monitoring
Limitations
- "Normal" is dataset-defined; concept drift triggers false alarms
- Class-imbalance evaluation is delicate
- Threshold calibration needs labelled anomalies