Wearable accelerometers and gyroscopes capture fine-grained behavioral signatures that can inadvertently reveal user identities, making privacy protection essential for healthcare applications. We present C-AAE, a lightweight compressive anonymizing autoencoder that performs on-device privacy filtering at the sensor edge. The core idea of C-AAE is to integrate two complementary privacy filters: a learned, sensor-specific anonymization module, the Anonymizing AutoEncoder (AAE), and a learning-free, generic anonymization module, Adaptive Differential Pulse-Code Modulation (ADPCM). The AAE locally learns to suppress identity cues while preserving activity-relevant representations, whereas ADPCM provides training-free anonymization through compression, further masking residual identity information and reducing communication cost. Experiments on the MotionSense and PAMAP2 datasets show that C-AAE cuts user re-identification F1 scores by 10-15 percentage points relative to AAE alone, while keeping activity-recognition F1 within 5 percentage points of the unprotected baseline. Implementation on a small-scale edge device (ESP32-WROOM-32) demonstrates real-time performance with markedly lower memory usage, latency, and power consumption. Unlike differential-privacy mechanisms that rely on randomized noise, C-AAE offers a complementary, representation-level approach, enabling practical and resource-efficient on-device anonymization that remains compatible with formal DP frameworks for hybrid deployment on edge healthcare devices.
Fujimoto et al. (Mon,) studied this question.