Resource Bounded Secure Goal Obfuscation

An AI system that is operating in an adversarial environment should be able to provide safeguards for its internal information. In other words, an adversary should not be able to perform diagnosis on the AI system’s internal information based on the resulting observations during plan execution. The system should be able to produce a plan that achieves the desired objective while minimizing leakage about its internal information. In this paper, we present an approach that allows an AI agent to securely obfuscate its true goal (i.e., agent’s internal information) for as long as possible using a subset of candidate goals. By making all the candidate goals equally likely for as long as possible, the agent’s true goal is kept secured. The AI agent may have to incur an additional cost to reach its true goal, but this cost buys the obfuscation guarantee. Given a larger resource budget, greater obfuscation is possible. We provide empirical evaluations of our approach using IPC domains collecting key metrics to show its feasibility.