Exploiting Partial Observability for Optimal Deception