In order for robots to coordinate effectively, they must communicate and share information. This communication can be attacked by malicious robots who alter the information they send and receive to prevent successful coordination — for example via spoofing attacks, where a robot transmits under multiple false identities. In the REACT Lab, we develop methods that provide resilience to multi-robot systems, allowing them to function and coordinate in the presence of malicious robots, including using Wi-Fi communication as a sensor to validate the information robots send. We apply these methods to persistent surveillance, flocking, distributed detection, and consensus, among other tasks.
Multi-robot adversarial resilience using control barrier functions
We design a controller that provides a multi-robot team with a communication network that is guaranteed to be resilient to adversaries, while maintaining that resilience as the team navigates an environment, satisfies other objectives, and avoids collisions. Modeling communication as a function of distance lets us treat resilience as a physical constraint, and we analyze when it conflicts with safety constraints — for instance when robots split around large obstacles or navigate narrow corridors. An alternative controller treats resilience as a soft constraint that can be momentarily sacrificed for other goals. This work was nominated for the Best Paper Award at RSS 2022.
Exploiting trust for resilient hypothesis testing
Robots sense the occurrence of an event of interest and relay their hypothesis as a binary measurement to a centralized fusion center, which also extracts a trust value from each transmitting robot. We develop two algorithms that use this information to improve the probability of arriving at the correct hypothesis: a Two-Stage Approach (2SA) that first decides which robots to trust and then fuses the trusted measurements, and an Adversarial Generalized Likelihood Ratio Test (A-GLRT) that jointly estimates trustworthiness and the true hypothesis. Applied to a crowdsensing scenario under a spoofing attack, both methods accurately predict traffic conditions despite a large number of spoofed robots.
Crowd vetting and dynamic crowd vetting
These works determine which robots in a network are malicious and which can be trusted, so malicious information can be filtered out. Using sensing over wireless channels, our FindSpoofedRobots algorithm utilizes neighboring opinions to speed up the validation process. In dynamic settings, robots performing random walks gather trust observations as they interact; by leveraging second-hand information from trusted neighbors, the time required to correctly estimate the trustworthiness of all robots remains constant as the team scales up, compared to logarithmic growth without it.
Adaptive malicious robot detection and local resilience
In persistent surveillance tasks, robot teams must maintain desired proportions across sites while detecting malicious robots that should not be counted — even when the network is dynamic, robots enter or leave at any time, and legitimate robots can turn malicious. We also analyze communication networks to find vulnerable areas where malicious information could spread quickly, and develop control laws that provide local resilience around a computed Vulnerable Set of high-influence robots, minimizing reconfiguration for the rest of the team.