CLIP-guided Diffusion Model for Backdoor Generation in Sensor-based Human Activity Recognition

The paper introduces the IMU-DM-CLIP, a backdoor training technique utilizing a diffusion model to facilitate trigger-based attacks on human activity recognition (HAR) systems that rely on Inertial Measurement Unit (IMU) sensors. The approach demonstrates effectiveness with a minimal backdoor injection rate of 10%, suggesting significant implications for the security of HAR models in IoT and wearable device applications. This research highlights the vulnerabilities in synthetic data generation for training HAR models, necessitating enhanced security measures for practitioners in the field.