Extracting Speech from Motion-Sensitive Sensors.

The increasing presence of wireless sensor networks and the blanket re-use of the resulting data volumes by AI-based systems raises pressing ethical questions about the impact of these technologies on our society. One of the commonly used technologies are Smart Phones and similar mobile communication devices which attract people to improve their quality of life. These devices are equipped with rich sensors that provide an advanced and comprehensive user experience. However, it is a well known problem that the presence of numerous sensors is of major concern to the privacy of users and their social environment. Specifically previous studies already revealed that motion-sensitive sensors actually react to human speech. In this regards Deep Neural Networks (DNN) proved very successful to model high-level abstractions in data. Our main focus is highlighting (i) the potential risks related to these sensors leaking private information about speech and (ii) the ethical implications of advances in (deep) machine learning as a threat to privacy. In this paper we showcase a simple attack in which collected data from accelerometer and Vibration Energy Harvester (VEH) sensors can be used to eavesdrop on speech. We propose a multistage stacked auto-encoder model that learns the distinctive time and frequency characteristics independently without user interaction. We demonstrate the efficiency of our model with poor quality data and a very low sampling rate. We investigated three classification tasks: gender identification (i), hotwords detection (ii), and (iii) recognition of simple phrases selected from a previously well investigated dataset. Our experiments demonstrate the efficiency of our model and confirm that motion-sensitive sensors are a rich source of personal data, from which highly sensitive and private information about people in close proximity to the sensor emerges.