Sensors fingerprints using Machine Learning: a case study on dam monitoring systems

This paper presents a comprehensive approach to getting sensors' fingerprints (i.e., identifiers based on physical properties) using machine learning classifiers. The object of study is a dam's structure health monitoring system, which constitutes a suitable tool to protect these critical infrastructures against hazard events. Since cybersecurity is a vital requirement in these systems, the correct identification of each component (including sensors) constitutes a basic premise to provide security mechanisms such as authentication and access control. We develop our study in both theoretical and practical aspects. After presenting the conjectures that basis our strategy, we implement an experiment that simulates conditions observed in a real dam's structure health monitoring system. We evaluate our experiment by testing four different machine learning classifiers. Our contribution portrays the hypothesis that the physical characteristics of piezoelectric sensors and the environment in which they are inserted are propagated through the captured signal. The results indicate an accuracy of over 90% in the correct sensors identification and attest our approach as a promising cybersecurity solution alternative.