Adversarial Causal Deception Scenarios: Preliminary Modeling and Policy Formation

Abstract

As autonomous systems become increasingly integrated into society, they may be presented with falsehoods and adversarial deception that can harmfully skew their perception of the state of their environment. Still tasked with making decisions in these contexts, it is thus important for systems to be aware of and plan around potential deception for optimal decision-making. This is inherently a causal problem given that deception often masquerades inaction as action, like a phishing attempt pushing urgency by spoofing a need that is not real. Agents successfully navigating these adversarial causal deception scenarios must understand what acts can truly change the state, and where misinformation is merely advertising that it has changed. This paper provides a causal framework for considering the portions of the state that are vulnerable to action and misinformation (an Adversarial Causal Decision Network (ACDN)), and outlines a planning process (Adversarial Causal Expectimax Search (ACES)) to avoid adversarial deception attempts in pursuit of the agent's purpose.