We design the first game-theoretic efficient machine unlearning evaluation metric with provable properties.