In slope stability prediction, challenges are posed by difficult sample acquisition, small data volume, and strong subjectivity in parameter selection. Traditional models are often afflicted by insufficient accuracy and overfitting. To address these problems, a CGAN-NSGA-II-BP model is proposed, in which a conditional generative adversarial network (CGAN) and the non-dominated sorting genetic algorithm II (NSGA-II) are integrated. First, constraint conditions and a penalty mechanism are introduced into the CGAN. The generated conditional variables are then filtered by the simplified Bishop method, by which the rationality and effectiveness of the augmented samples are physically guaranteed. Second, optimal hyperparameters of the BP neural network under the mixed dataset are automatically searched by NSGA-II, whereby overfitting is effectively alleviated and generalization performance is improved. The improved model is compared with multiple existing models, and an engineering case verification is carried out. It is demonstrated by the results that high consistency in overall distribution pattern, numerical range, and inter-variable correlations is maintained between the augmented and original data. A JS divergence of only 0.1462 is obtained, a Frobenius norm difference of only 0.1345 is observed for the correlation coefficient matrices, and all coefficients of determination R2 of the conditional regression are found to exceed 0.9. Under the experimental conditions, an accuracy of 96.08%, a precision of 96.06%, a recall of 97.33%, an F1 score of 0.9667, and a mean AUC of 0.9894 are achieved by the CGAN-NSGA-II-BP model, and its overall performance is demonstrated to be significantly superior to those of the compared models. The effectiveness and practicability of the model are verified by the engineering case, and a reliable new method is provided for slope stability prediction under small-sample data scenarios.