PIs: Arnab Bhattacharya, Ted Fujimoto, Himanshu Sharma

The recently passed U.S. infrastructure bill seeks to modernize the country’s cyber-defenses in critical infrastructures and strengthen capabilities to secure, respond, and coordinate operations in real-time at targeted infrastructure assets. As outlined in the bill, a key research area of interest is improving resilient operations via threat detection and coordinated automation in transportation systems with intelligent sensor infrastructure.

This project is developing a planning-based algorithmic approach and computational pipeline for online adversarial threat inference and mitigation for a fleet of autonomous vehicles (AVs), subject to opportunistic exploitation of their navigation and communication (NAVCOM) sensors. The goal of our project is to demonstrate the efficacy of the proposed learning approach in identifying adversarial goals, localizing compromised sensors, and steering AV fleets toward their desired task objectives via multi-agent coordination strategies in online operational settings.

We consider a problem setting where an AV fleet equipped with multiple heterogenous sensors is required to execute a go-to-goal navigation task. An adversary’s mission is to exploit and direct the AVs to an undesired location by perturbing a set of sensors whose NAVCOM measurements lie within the system’s uncertainty bounds. A key novelty of our algorithmic approach is the capability to rapidly infer shifts in adversarial intent from system-state trajectories affected by adversarial strategies that can be either optimal or sub-optimal. Here, intent refers to a collection of multi-step attack sequences used to reach a target adversarial goal that leads to undesirable operational behavior and system performance.

Our learning approach enables defensive systems to infer an adversary’s intent online (by actively interacting with a simulated environment model), identify the set of compromised sensors during simulated AV spoofing and jamming exploits, and guide the AV fleet toward their desired objective using multi-agent coordination strategies.

The major components of this project include the following:

• a sensor-fusion module to aggregate online measurements from different sensors to estimate the system state
• a threat inference engine that uses Bayesian sequential inverse search and planning algorithms to dynamically infer adversarial intent from compromised AV sensor data
• a defense-resource optimization engine that identifies the uncompromised sensors—these measurements can be used to determine an effective multi-agent coordination strategy to steer the fleet toward their desired objective and guarantee safety assurance
• a high-fidelity domain-guided AV simulation environment for testing the threat inference and defense optimization engines over diverse adversarial and system uncertainty scenarios.