Abstract: An adaptive key region sampling method is proposed for evaluating structural reliability by establishing a deep neural network surrogate model. This adaptive key region sampling method combines distance information and probability distribution information, allowing for the consideration of the significant impact of samples near the limit state surface, as well as the influence of sampling points following the global probability distribution, striking a balance between local exploration and global exploitation. To avoid the reduction in sampling efficiency due to the clustering of sampling points, a candidate point removal rule is proposed. Considering the characteristics of deep neural network surrogate model, a uniform Latin hypercube sampling experimental design is adopted for initialization, along with convergence criteria that take into account the fluctuation characteristics of deep neural network predictions, ensuring the robustness of the proposed algorithm's convergence. Validation through three numerical examples demonstrates that the method presented in this paper exhibits significant advantages in terms of accuracy and efficiency.