Slope reliability analysis is a critical aspect of geotechnical engineering, particularly under conditions of rainfall infiltration, where the spatial variability of soil parameters can significantly affect the reliability of slopes. Traditional methods like Monte Carlo simulation are often computationally intensive, severely challenging the design of cutting slopes considering the spatial variability of multiple soil parameters. To address this challenge, this study proposes a convolutional neural network (CNN)‐based surrogate model to efficiently assess the reliability of unsaturated soil slopes. The CNN model is trained to establish an implicit relationship between the random field inputs of soil parameters and the corresponding slope stability outcomes, enabling rapid calculation of the probability of failure (Pf) under varying conditions. The results indicate that as rainfall intensity increases, the Pf rises. For the same slope cutting distance, a greater slope cutting angle leads to a higher Pf. Similarly, for the same slope cutting angle, increasing the slope cutting distance results in a higher Pf; and the impact of slope cutting distance on slope reliability is more significant than that of slope cutting angle. Additionally, for various rainfall conditions and slope cutting scenarios, the CNN‐based surrogate model is integrated into the full probability reliability design method, and a design response surface is used to establish the relationship between design variables and reliability responses. It is found that the proposed approach can efficiently evaluate the reliability of all design schemes. A strategy for determining the optimal slope cutting scheme is finally provided as practical guidance to meet the target reliability.