This paper proposes a nonlinear structural model updating approach based on the improved Generative Adversarial Net work （GAN）. In the improved GAN， the ability of the network discriminator to learn the response characteristics of nonlinear structures is enhanced by using the surrogate models. To avoid the problem that the gradient disappearance problem existing in traditional GAN， information exchange between network layers is strengthened by means of skip connection and dense connection， and a combined objective function is added to the improved GAN， whose aim is to construct the mapping relationship between structural responses and model parameters of a nonlinear structure to realize the network training. In nonlinear model updating， structural dynamic responses are considered as input of the network. After training， the constructed GAN network can predict the optimal value of nonlinear model parameters by learning the characteristics of input data， and the nonlinear model updating can be performed. To validate the feasibility of the proposed method， a 12-story reinforced concrete frame structure under earthquake excitations is conducted as numerical simulation， and the accuracy of the proposed method is further verified by comparing the results of nonlinear model modification based on convolutional neural network. Finally， the proposed method is applied to a nonlinear model updating of a cantilever beam shaking table experiment under earthquake excitation， and the same good updating results are achieved.