Application Specific Bounds on Detection Cost using Game Theory

In some speaker recognition applications the rate of impostor claims will not be known in advance and may in fact be under the control of the impostors. This prevents the implementers of the recognition system from simply choosing an operating point to minimize the detection cost function. We present an analysis of the adversarial game that results from both attacker and defender being able to choose a strategy and show that the system implementers can choose an operating point chosen that in game-theoretic optimal performance and bounds on maximum cost per trial to the implementer. The result also predicts the optimum impostor claim rate that will maximize the impostor´s rewards