Privacy-preserving and communication-compressed multimodal activity recognition for mobile terminal sensing information

Abstract

Continuous activity recognition on mobile terminals can draw on inertial, location, Wi-Fi, and Bluetooth signals, yet rawstream upload exposes behavior patterns and adds communication, latency, and battery cost. PPCC-Fuse is proposed as a privacy-preserving and communication-compressed multimodal activity recognition framework for mobile terminal sensing information. On the terminal side, local features are extracted and a Context-Aware Risk-Rate Controller (CRRC) coordinates feature selection, privacy-aware perturbation, adaptive quantization, and delta-based upload. The edge server then decodes compressed feature packets and applies missing-modality robust fusion. A real deployment with 32 users, four Android terminal types, mixed Wi-Fi/4G/5G links, and 164.3 h of labeled sensing data produced 236,592 valid windows. Compared with local feature upload, PPCC-Fuse reduces upload volume from 14.62 to 2.79 KB/min and useridentification attack accuracy from 64.8% to 27.9%, while Macro-F1 changes from 92.1% to 90.8%. Compared with raw multimodal upload, it reduces mean latency from 238.6 to 65.4 ms and calibrated terminal energy from 24.8 to 11.6 J/min. Seven-day MQTT QoS 0 logs show an overall packet loss rate of 0.53% for continuous sensing packets. The results indicate that coordinated terminal-side privacy control and communication compression offer a practical trade-off for mobile activity recognition under edge deployment constraints.