SpectrumAudit identifies sensor biases in wearable activity recognition models
September 25, 2026
SpectrumAudit provides a label-free auditing method for wearable HAR models by applying phase-randomized stimuli to identify sensor offsets. Testing across 27 models showed accuracy losses between 2.87 and 40.83 percentage points, proving that models often exploit DC sensor biases rather than temporal patterns.
HOW THIS AFFECTS YOU
●
builderYou should audit your wearable models for sensor offset exploitation to ensure robustness across different hardware.
●
researcherThis offers a new method for label-free auditing of temporal models without needing ground-truth labels.