Deception-Resistant AI

PI: Zachary Sunberg

• Optimal control and MDP-based deep reinforcement learning (RL) do not reason about active deception.
• Game theory provides mathematical tools and algorithms (e.g. DeepStack AI for Poker).
• Our task: extend these game-theory concepts to autonomous aerospace systems ("drone poker").

How can AI systems resist deception by other agents and uncertainty in nature?

Funding: ONR Science of Autonomy "DECODE-AI" Project

Project 1: Combining robotic and game-theoretic uncertainty representations

Project 2: Neural networks for multi-modal continuous policies.

Particle filters from robotics

Information sets from game theory

[Becker & Sunberg AAMAS '25 Ext. Abs.]

Example: Drone defense - how do we calculate deceptive trajectories and counter them?

Resulting stochastic discrete policies

Traditional Gaussian policies are unimodal and exploitable

Flow-Matching magnetic mirror descent produces multimodal unexplotable policies

[Becker & Sunberg AAMAS '27 (Under Review)]