Membership Inference Against Self-supervised IMU Sensing Applications

Deep learning has revolutionized the use of inertial measurement unit (IMU) sensors in mobile applications, such as human activity recognition. Building on the success of pre-trained models across various domains, recent studies have increasingly adopted self-supervised learning (SSL) for a range of sensing tasks. While these SSL approaches improve generalization and reduce labeling requirements, their privacy implications have received limited attention. This paper addresses this gap by examining IMU data privacy during pre-training through membership inference. Our work serves two important purposes: First, it enables data owners to verify if their data was used without permission in encoder pre-training. Second, it demonstrates how adversaries might compromise sensitive human sensing data used in pre-training. To enhance the practicality of membership inference on unlabeled IMU sensing data across different SSL algorithms, we introduce an activity labeling module and a novel perturbation strategy to exploit encoder overfitting characteristics on training data. When an encoder over-fits, it memorizes training data rather than learning generalizable patterns. Therefore, when comparing the original data to the perturbed version, the encoder generates more distinct feature vectors for samples from its training set than for samples it has never seen before. We evaluate our membership inference methods on two mainstream SSL methods across multiple datasets, demonstrating that our method can achieve relatively high precision and recall at low false positive rates.