Yerin Kim

In Markov Decision Processes, an observer can infer an agent’s underlying objectives from its behavior using inverse reinforcement learning (IRL). We design policy regularizers that steer these inferences toward false rewards while preserving task performance.

Key result. We derive bounds on accumulated reward loss as a function of regularization strength. In the paper’s network-defense simulations, the three deception strategies induced false beliefs in three different IRL algorithms, while retaining at least 98% of optimal non-deceptive reward.

How can a controller introduce deceptive perturbations into its trajectories while keeping those perturbations difficult to detect? We augment a quadratic control cost with a likelihood-ratio term, derive the optimal controller, and analyze how the probability of detection varies with deception intensity.

Key result. We develop methods for selecting deception intensity to keep the probability of detection below a prescribed tolerance ε at each time step, using sampling-based probability estimates and analytical bounds:

P(log L_t^* ≥ log U) ≤ ε, for every t = 1, …, T.