TrustTiny-HAR: Selective, Open-Set, and Calibrated Activity Recognition on Microcontrollers

Human activity recognition (HAR) increasingly executes on microcontrollers to satisfy privacy, latency, and energy constraints, yet most deployed TinyML models remain closed set and overconfident when faced with unseen behaviors, placement shifts, or sensor anomalies. We presentTrustTiny-HAR, a resource-efficient pipeline that unifies an int8 time-series backbone with a distilled trust head and a lightweight streaming conformal layer. The unified head produces calibrated class probabilities and a compact out-of-distribution (OOD) surrogate score learned from teacher signals (energy of logits and prototype/ Mahalanobis distance), while the conformal layer maintains tiny on-device quantiles to enforce coverage-controlled acceptance thresholds. When evidence is insufficient, the device abstains and may optionally escalate using a 32–64D feature sketch, avoiding raw-signal transmission. Evaluated under leave-one-subject-out protocols on UCI HAR, PAMAP2, and WISDM with multiple OOD stressors (unseen classes, placement shift, injected anomalies), TrustTiny-HAR matches strong int8 baselines on indistribution macro-F1 (e.g., 93.8 on UCI HAR) while substantially improving calibration (ECE 2.4%) and OOD robustness (AUROC 0.96 for unseen, FPR@95 0.22 for placement shift). As a selective classifier, it attains the desired 90% coverage with conservative risk (e.g., 6.1%) from a buffer of only 128–512 windows. System profiling on nRF52840/STM32L4 showshttps://github.com/Ism-ail11/TrustTiny-HAR.