In order to achieve the higher security level of power system in context of cyber-physical fusion, the detailed procedures of attacks and defenses are studied in this paper based on the three-level framework of game theory. The optimal defense strategy and the minimal load loss expectation under the limited defense resources are also investigated. Firstly, the optimal defense strategy is modeled in the upper defender model to reduce the vulnerability of power grid. Then, the middle attacker model is established to maximize the attack effects. In the lower-level, the optimal dispatch is used to reduce the attack loss. To solve the problem efficiently, mixed integer programming method combined with traversal thought based genetic algorithm is employed to calculate the optimal results. The feasibility of the proposed approach is validated through a benchmark case study. This work is supported by National Natural Science Foundation of China (No. 51437003).