Typical Composition of a Cyber System
Typical Composition of a Cyber-physical System and a Sensing-and-actuation Feedback Loop
Cyber systems can simply be characterized as "connected systems controlled by software." In addition to control and in/output interface(s) for direct human control, these systems are typically equipped with communication interface(s), through which they can interact or cooperate with other systems. The most notable characteristic is the software, which enables cyber systems to work (semi-)autonomously without human inputs.
In comparison, cyber-physical systems (CPS) are "cyber systems extended to the real world." A typical CPS comprises all the characteristics that a typical cyber system has, and in addition, it is equipped with communication interfaces for the real world: sensors and actuators. Both are one-directional interfaces but with the opposite directions. Sensors transfer information from the real world to the system, whereas actuators realizes the decision of the system in the real world. With these two in/output interfaces toward the real world, a feedback loop of sensing and actuation is formed, and such a CPS is referred to as "sensing and actuation system."
The figure collection below lists representative examples of cyber-physical systems: autonomous vehicles, drones, IoT devices, industrial control systems, and medical devices. As can be seen in the figures, various types of sensors are mounted on them.
Representative Examples of Cyber-physical Systems and Mounted Sensors
Sensing-and-actuation Feedback under Benign Circumstance
Sensing-and-actuation Feedback under Sensor Attack
Sensor attacks adversarially exploit the sensing-and-actuation feedback. Under benign circumstances, environmental changes are well received by the sensors and a CPS accordingly reacts to the changes with the actuators mounted on it. However, under sensor attacks, the attacker exposes sensors to malicious signals which convey falsifying information on the surroundings. These malicious signals are generally optimized so that they can overwhelm legitimate signals carrying true information on the surroundings. Being sensors are the sole or major source of information on the real world, a compromised CPS erroneously reacts to such falsifying information given by the attacker.
Unlike cyber systems which are indirectly coupled with the real world, such an erroneous behavior of CPSs directly affects humans and properties. For example, a compromised autonomous vehicle may hit pedestrians or collide with other vehicles; both would lead to direct damage to humans or properties. Note that all the threats against cyber systems may also work against CPSs because a CPS is also a cyber system.
The number of attack vectors against a specific sensor is highly related to how complicated the sensor is. Generally, the more complicated a sensor is, the more ways there are to compromise it. For example, it is quite limited to affect a simple camera sensor by a sensor attack. Just exposing the camera to intense light to blind it would be feasible. On the other hand, however, a complex vision sensor, e.g. Mobileye, with automatic object detecting and locating capability would additionally likely to be affected by attacks like physical adversarial examples.