ADVERSARIAL OPTIMIZATION METHOD FOR TRAINING PROCESS OF GENERATIVE ADVERSARIAL NEURAL NETWORK

摘要

An adversarial optimization method for a training process of a generative adversarial neural network. In a generator G, an optimal transmission problem is converted into a solution to an elliptic-type Monge-Ampere partial differential equation (MAPDE), and in order to solve the MAPDE of n (n>3) dimensions, a Neumann boundary condition is improved and the discretization of the MAPDE is extended, so as to obtain optimal mapping between the generator and a discriminator, thereby constituting an adversarial network MAGAN. According to the method, during the process of training a defense network, by means of overcoming a loss function of optimal mapping, the defense network can obtain the maximum distance between two measures, obtain filtered safe samples, and successfully establish an effective attack method for GANs, such that the precision is improved by 5.3%. In addition, an MAGAN can be stably trained without adjusting hyper-parameters, thereby greatly improving the accuracy of an unmanned driving target classification and identification system.