WearGuard: Defeating Adversarial Attacks in Wearable AI-based Activity Recognition Systems

Inertial Measurement Units (IMUs) on wearables have revolutionized HAR in healthcare, fitness, and smart environments. In recent years, a large body of work has adopted deep neural networks as the backbone for IMU-based HAR, achieving high classification accuracy across various benchmarks. However, these deep neural networks are vulnerable to imperceptible adversarial perturbations, leading to misclassifications and posing severe risks in safety-critical applications such as elderly monitoring, rehabilitation, and industrial safety. This paper introduces WearGuard, a comprehensive defense framework designed to enhance the adversarial robustness of IMU-based HAR systems. The framework combines a signal-level module that reconstructs clean representations from corrupted inputs using a lightweight autoencoder and multi-teacher supervision, and a model-level module that maintains temporal consistency and preserves semantic relationships in the learned embeddings. Extensive experiments across multiple datasets and neural architectures demonstrate the effectiveness of our approach against four major adversarial attacks. WearGuard recovers over 85% of clean accuracy under different adversarial conditions. © 2025 Copyright held by the owner/author(s).