Differential Privacy with Weighted ∊ for Privacy-Preservation in Human Activity Recognition

Many services based on human activity recognition (HAR) have been developed; however, user activity data include a large amount of private information. Although privacy protection is important in activity recognition, it has not been sufficiently explored. Therefore, we propose a privacy-preserving mechanism for HAR services that uses differential privacy. The proposed method reduces the user recognition accuracy to a level that satisfies the privacy requirements by adding weighted noise to the features in the learning model construction and then improves the activity recognition accuracy (service usefulness). The results indicate that when the privacy requirement is defined as less than the probability of a user being identified by chance, the proposed method improves the activity recognition accuracy by approximately 10 % compared to the conventional method.