Understanding the behavior of drone swarms
Parenting a toddler is like flying a drone. It’s easier when there is just one of them; when there are more, it’s difficult to predict what they will do next – and there’s a chance one could go rogue at any time.
While parents may not be able to predict toddlers’ actions, designers of drone swarms could soon understand the correct behavior of a group of unmanned aerial vehicles (UAVs), alerting operators if any are behaving irregularly or have potentially been hacked. Sibin Mohan, Research Assistant Professor in the Department of Computer Science and the Information Trust Institute, has a new project titled, “Characterizing Behavior and Anomaly Detection in Distributed Unmanned Vehicles,” which will outline what constitutes normal behavior for these types of systems.
“If we can detect that the system, or parts of a system, is not functioning correctly, then we can fix it,” said Mohan. “If you can’t detect an error, then the operator may not know that an action has to be taken, much less what actions to take.”
In instances where only one UAV is used, the controller largely knows what the vehicle is supposed to do. However, when many UAVs interact with one another, then such interactions can change their behavior.
“If multiple vehicles are communicating with each other and something looks odd, it could be for a number of reasons,” Mohan explained. “Maybe they are too far apart, the environment is causing abnormal behavior, someone has hacked one or more of them, or some unexpected group behavior has emerged. You have to know what the correct or expected behavior is to know something is odd.”
One of the reasons UAV and ground rover swarms are so vulnerable to attacks is because they communicate with each other through wireless communication. This system is unreliable and easy to jam, which makes it even more important to know how group of vehicles should operate.
This project is a three-year collaboration with the Boeing Company. Mohan, along with graduate and undergraduate students, will be building test platforms to study and characterize normal vehicle behavior, using machine learning techniques.
“We haven’t begun building the platforms yet,” said Mohan. “As part of the collaboration we will demonstrate our results on a realistic platform so it’s not purely theoretical research.”