Smart-Glasses IMU Streams as Cybersecurity Side Channels

Abstract

Smart glasses continuously collect inertial signals for head tracking, stabilization, interaction, and context awareness. From a cybersecurity perspective, these apparently operational signals form a behavioral side channel: they may allow an untrusted application or compromised sensing pipeline to infer what a user is doing even when cameras and microphones are unavailable. Recent AR/VR security studies show that headsets and associated physical channels can leak hand gestures, keystrokes, voice commands, speech, and application behavior. This paper evaluates the strength of the IMU side channel by benchmarking deep learning approaches on inertial measurement unit (IMU) streams captured by Meta Aria smart glasses. We focus on the full 33-class experimental setting and compare LSTM, GRU, 1D CNN, CNN-LSTM, CNN-GRU, a classification Transformer, and a Two-Stream Transformer Network (TTN). The best models achieve high and closely grouped performance: TTN reaches 92.13% accuracy and 0.9215 macro-F1, while the 1D CNN reaches 91.42% accuracy and 0.9143 macro-F1. Mutual-information and inter-model-agreement analyses further show that the strongest models learn similar leakage structures. The key security finding is that accurate behavioral inference does not require visual data or a heavyweight architecture; compact models can exploit non-visual IMU telemetry. These results motivate treating smart-glasses motion streams as privacy-sensitive attack surfaces in AR/VR platforms.