Practical Privacy-Preserving Behaviour-Based Authentication on the Edge
Andraž Krašovec, Veljko Pejović · 2022
Sensing-enabled Internet of Things (IoT) devices provide an attractive source of data that can be used for seamless authentication of people located within the reach of these sensors. However, current behaviour-biometric solutions often require that the data is centrally stored and processed. This may pose privacy risks, as the data may be used not only for identifying a user, but also for inferring a user’s activities within an IoT environment. The risk can be alleviated by shifting the processing to the edge, so that only selected, obfuscated data is sent to the cloud. Previous work proposes a conceptually elegant privacy-preserving distributed deep learning inference method for user identification. In this work we attempt to actually implement such a method distributed over edge devices and the cloud, and uncover significant challenges related to hardware and system limitations of the distributed setup. We then propose amelioration strategies for privacy-preserving distributed deep learning training. Nevertheless, these strategies only partly address the identified issues, calling for further work in bringing distributed sensor-based behaviour authentication to reality.