Generalization and suppression are important technical approaches for k-anonymity, which protect respondents' identity when releasing microdata. However, although those solutions are very efficient in terms of the security of sensitive private information, they are known to be defective. They may cause some deviations between the published dataset and the original dataset, the authenticity, readability, and applicability can be reduced to the tuples in the released dataset. In this paper, we proposed a new and powerful improved method, called δ-generalization, to solve these problems. We show, thorough both theoretical analysis and experiment evaluation, that δ-generalization can significantly enhance the accuracy of the published dataset on the premise of guaranteeing the equal security levels compared with the conventional generalization. Extensive experiments with real data confirm that δ-generalization is practical and can be implemented efficiently.