Deep neural networks are known to be vulnerable to small adversarial perturbations to the inputs. Empirical methods such as adversarial training can defend a particular class of attacks but can still be `broken' by stronger attacks. Another approach to constructing the Lipschitz network is certified robust to perturbations, but its expressive ability is not satisfied. To combine the advantages of the two methods, we design a novel two-step model that can fit the training data accurately and preserve the local Lipschitz property simultaneously. We first train ResNet with a regularizer from optimal transport theory and obtain a discrete optimal transport map from data to its features. As the optimal transport map has regularity property, we interpolate the map by solving the convex integration problem to guarantee the local Lipschitz property of the interpolation. Numerical experiments show that this model can outperform other robust models on diverse datasets well-trained by a ResNet with latent dimensionality unchanged.