Non-intrusive Continuous User Identification from Activity Acoustic Signatures

Recent advances in acoustic sensing have led to the rapid proliferation in the deployment of microphones in smart-environments. This, in turn, has allowed the infrastructure to capture various types of ambient noise as a part of the contextual information. In this paper, we explore one such possibility of using acoustic signatures generated by human activities to distinguish individuals performing the activities continuously. We achieve this by developing an automated machine learning-based framework, trained on basic acoustic features, that can identify the users from the acoustic signatures generated by their activities. However, developing such a system has several associated challenges. This paper identifies those challenges and performs a preliminary evaluation of the system to determine the current approach's limitations and possible future directions.