In deep learning model, the training procedure can obtain appropriate weights by leveraging the information of gradient descent. However, the gradients between the weights may leak the original neural network training data if they are not well-designed. Zhu et al. (Advances in neural information processing systems, Curran Associates, Inc., [5]) as proposed a kind of attacking scenario named deep leakage from gradients. There were many mitigation strategies proposed in the literature, such as mixing Gaussian noises and model quantization. However, they usually somewhat lost classification accuracy during prevention process. This paper proposes a lightweight attachment model which can be used as an extension on the original classification model to prevent deep leakage from gradients with low information loss. This approach can complete the goal of mitigation on trying to minimize the changes of the original structure of the model and can be easily implemented on some existent deep learning classifiers.