Context-Aware Sensor Fusion for Securing Cyber-Physical Systems

The goal of this dissertation is to provide detection and estimation techniques in order to ensure the safety and security of modern Cyber-Physical Systems (CPS) even in the presence of arbitrary sensors faults and attacks. We leverage the fact that modern CPS are equipped with various sensors that provide redundant information about the system\u27s state. In such a setting, the system can limit its dependence on any individual sensor, thereby providing guarantees about its safety even in the presence of arbitrary faults and attacks. In order to address the problem of safety detection, we develop sensor fusion techniques that make use of the sensor redundancy available in modern CPS. First of all, we develop a multidimensional sensor fusion algorithm that outputs a bounded fusion set which is guaranteed to contain the true state even in the presence of attacks and faults. Furthermore, we provide two approaches for strengthening sensor fusion\u27s worst-case guarantees: 1) incorporating historical measurements as well as 2) analyzing sensor transmission schedules (e.g., in a time-triggered system using a shared bus) in order to minimize the attacker\u27s available information and impact on the system. In addition, we modify the sensor fusion algorithm in order to provide guarantees even when sensors might experience transient faults in addition to attacks. Finally, we develop an attack detection technique (also in the presence of transient faults) in order to discard attacked sensors. In addition to standard plant sensors, we note that modern CPS also have access to multiple environment sensors that provide information about the system\u27s context (e.g., a camera recognizing a nearby building). Since these context measurements are related to the system\u27s state, they can be used for estimation and detection purposes, similar to standard measurements. In this dissertation, we first develop a nominal context-aware filter (i.e., with no faults or attacks) for binary context measurements (e.g., a building detection). Finally, we develop a technique for incorporating context measurements into sensor fusion, thus providing guarantees about system safety even in cases where more than half of standard sensors might be under attack