Can Deep Learning Infer Age, Height, and Weight from Human Activity Sensor Data?

Sensor-based human activity recognition is increasingly used as a key task in many applications. However, collecting and processing data coming from the many sensors commonly available in personal devices can pose a threat to user privacy, especially when the signals are transmitted to the cloud. Assessing the extent to which deep learning can infer personal data, such as height, weight, and age, from motion data remains an open and delicate question. This study explores the performance of convolutional neural networks in inferring users’ personal features from a few seconds of accelerometer traces collected by their smartphones during simple movements. The performance achieved by convolutional models trained on a large data set of annotated 2-second acceleration traces belonging to 146 subjects performing well-defined activities confirms that personal information can be inferred with a prediction accuracy that threatens user privacy, although the performance strongly depends on the type of activity observed and on the quality of the training set. In particular, subject-out experiments show a limited generalization capability.