Researchers at the Harvard John A. Paulson School of Engineering and Applied Sciences, working with a multi-university team, have developed a framework aimed at improving the reliability and safety of connected autonomous systems by introducing a quantitative measure of trust between machines.
The study, led by Stephanie Gil, outlines the concept of “cy-trust,” defined as a numerical value representing how much one autonomous agent, such as a robot or vehicle, should rely on information received from another agent or data source. The framework is intended for use in cyber-physical systems, including ride-share fleets, automated trucking convoys, and smart infrastructure networks.
The authors argue that existing cybersecurity approaches, which typically focus on controlling system access, are insufficient for systems in which machines must continuously exchange and act on real-time information. In such environments, inaccurate or malicious data can influence collective behavior and lead to physical consequences, including traffic disruptions or safety risks.
The paper identifies several potential threats specific to multi-agent systems. These include manipulation of shared data, such as falsified traffic information, and adversarial behavior by individual agents, such as misreporting position or identity. The researchers note that these vulnerabilities can affect coordination in applications ranging from transportation to emergency response.
To address these risks, the proposed framework incorporates data validation mechanisms using onboard sensors and communication signals. Systems equipped with cameras, lidar, radar, and GPS could cross-check externally received data against locally observed conditions. Signal-processing techniques applied to wireless communications may also help verify the origin of transmitted information.
Within the framework, each agent assigns a trust score between zero and one to incoming data based on factors such as sensor input, network behavior, and prior interactions. These scores influence decision-making processes, allowing systems to discount or ignore inputs deemed unreliable. According to the authors, this approach could help prevent system-wide disruptions caused by compromised or deceptive agents.
Experimental work conducted by the research team includes simulations in which cooperative robots attempt to reach consensus while adversarial agents introduce false information. In these scenarios, the system evaluates message sources and adjusts trust levels over time, enabling it to identify and disregard malicious inputs while maintaining coordination among reliable agents.
The researchers state that broader adoption of such frameworks may depend on their integration into system design standards and regulatory approaches, particularly as autonomous and interconnected technologies expand into public environments.
