Networked control systems — autonomous vehicles, drones, factory floors — pull in vast amounts of sensor data, which makes them attractive targets for remote attackers who spoof measurements or hijack actuators.

We treat attack detection as a statistical change-detection problem and design dynamic watermarking: the controller injects a private, random probing signal and checks whether the system’s response has the statistics it should. If an attacker tampers with the loop, the watermark’s signature breaks and the attack is exposed.

Highlights

• A dynamic watermarking algorithm for finite Markov decision processes, with guarantees on detecting a class of attacks.
• Model-free change-point detection for mixing processes, with unknown pre- and post-change kernels.
• Game-theoretic control policies against bus-off attacks in CAN networks.

Related work appears on the Publications page and connects to the group’s broader security research.