CAIN: Counter Autonomy Inference Network

We will build a prototype (based on the results of Phase I studies) which is able to provide behavioral exploitation, adversarial force modeling, and network inference. CAIN will perform risk-aware experimentation on unaware or unwitting autonomous systems by simulating and predicting expected behavior. This capability will gather intelligence while being aware of how much additional risk is undertaken for each action. To predict adversarial behaviors, we will model the cyber-physical systems of adversarial platforms and impacts intervention strategies will have on adversarial autonomy behavior, providing an understanding of the adversary’s communication methods, strategies, and dependence. Under this capability, cyber-physical models will be built for adversarial forces that will then use Theory of Mind inference methods to predict adversarial reaction to intervention strategies. Furthermore, intelligent stopping criteria will be used to balance multiple objectives linked to the Theory of Mind (ToM) of an autonomous platform. Under this capability, intervention strategy impacts on adversaries will be predicted and will balance the trade-off between accuracy of classification and risk to discovering intent by adversaries to send a stop-signal to end the operation. The system developed under CAIN will have a two-part feedback system composed of an Intervention Decision Aid (IDA) and a Level of Autonomy (LoA) Classifier. The IDA and LoA classifier will both balance covert tactics with behavioral classification of autonomous platform architecture. Sensors will provide environmental and situational information that is fed into this two-part feedback system of CAIN. The IDA and LoA classifier represent the core work proposed under CAIN to ingest sensor information and classify the level of autonomy of a platform. AUTOSENSE utilizes a hypergame planner on a Markov decision process (MDP) to decide which intervention strategies to take that simultaneously limits the risk to discovering intent by an adversary and increases the accuracy of the LoA classifier.