Smartphone-Based Place Profiling in a Privacy-Preserving Manner

Places and their semantics are the primary keys for various mobile applications and location-based services. Currently, smartphones can easily access users' locations in terms of longitude and latitude readings with different levels of accuracy. On one hand, these raw readings provide little values or even ambiguity to those mobile applications due to their low resolution in indoor places. On the other hand, directly leaking these readings to the third party applications may violate users' location privacy. In this paper, we propose a novel place profiling method to resolve these issues using common anonymized smartphone sensor data including accelerometer, WiFi, Bluetooth, light and sound. Here the anonymization refers to the location identifiable information in the sensor data, e.g. the mac addresses of WiFi access points, is replaced by meaningless symbols. Based on such anonymized sensor data, the proposed place profiling method is able to recognize different places with high level of accuracy. Our place profiling method includes a feature engineering step to extract 19 different features to profile each place and a unsupervised clustering method based on either Gaussian Mixture Model or K-Means. Experiments show that when the number of places is small, e.g. from 3 to 5, the achieved clustering accuracy ranges from 92.5% to 85.2%. Even when the total number of places is 9, the clustering accuracy is 71.6%. This study suggests that it is possible to provide location-based services using the anonymized sensor data even without the exact location information, and also it is safe to combine different users' data without the concern of leaking their location privacy to further improve the place profiling.