Protecting Safety- and Mission-Critical Information


Efforts focus on protecting safety- and mission-critical information by assuring privacy in computations, communications, and mission execution. With privacy-preserving computing, autonomous agents can assure the security of their data not just during communication, but also during computation. Given resource constraints and bandwidth limitations required to integrate the necessary cryptographic methods within the design envelope developed in other research areas (e.g., dwell-times and latency issues), these methods must be performant and robust against a strong attacker model. To protect communication events, we consider differential privacy, a statistical notion of privacy that makes it unlikely for an adversary to learn any meaningful information about agents, enforced by adding noise to sensitive data (or functions of sensitive data) before they are shared. However, operating in adversarial settings imposes the additional challenge of asynchrony in agent communications which, in conjunction with stochasticity, introduces fundamental analysis and design challenges to private communication. Conventional privacy analyses quantify privacy's impact using statistical notions such as variance or entropy, but the proposed efforts focus on privacy in networks of agents with hybrid dynamics, and we will quantify these impacts and investigate new mechanisms (e.g., using concepts from stochastic geometry) to reduce them. To protect information that reveals mission objectives and modes of operation, we will explore the use of opacity. Opacity characterizes a system's ability to conceal its "secret" information from being inferred by outside observers, and is encoded by ensuring that no observer can confidently determine an agent's mode with probability beyond some threshold. Our efforts are based on the key insight that the observers's belief update dynamics can be characterized as a discrete-time switched system whose switching signals are the observed actions. This observation offers a bridge between the widely used family of models, namely POMDPs, and concepts from control theory (e.g., barrier certificates). We leverage this novel approach to the synthesis of control protocols to satisfy mission or safety requirements while enforcing additional requirements on information flow patterns, motivated by privacy and/or security, as well as other applications (e.g., active learning) that also naturally admit formulations based on POMDPs.


Protecting Safety- and Mission-Critical Information Publications