Abstract: To address the security and stability problem of power systems in the restoration process after cyber-attacks while coping with environmental uncertainties, a surrogate gradient-based deep reinforcement learning for power system post-contingency safety control strategy against cyber-attacks is proposed in this paper. First, the cyber-attack models against information system data and functions are established, and the system security control model is constructed. The evolution process of system events under cyber-attacks is analyzed. Second, the Markov decision process of the security control strategy is defined under the framework of deep reinforcement learning. Then, a surrogate gradient-based deep reinforcement learning algorithm is designed, where the agent population is generated by perturbing the agent parameters, and the weighted average of the fitness values corresponding to each perturbation is used as the surrogate gradient. Finally, the effectiveness and superiority of the proposed method are verified on the IEEE 39-bus system.