A multi-model cooperative defense method facing deep learning antagonism attack comprises the following steps of: 1) performing unified modeling based on a gradient attack to provide a [Rho]-loss model; 2) according to design of a unified model, for an countering attack of a target model fpre(x), according to a generation result of a countering sample, classifying a basic expression form of an attack into two classes; 3) analyzing the parameters of the model, performing parameter optimization of the [Rho]-loss model and search step length optimization of a disturbance solution model for the countering sample; and 4) for the mystique of a black box attack, designing an experiment based on an adaboost concept, generating a plurality of different types of substitution models, used to achievethe same task, for integration, designing a multi-model cooperative defense method with high defense capability through an attack training generator of an integration model with high defense capability, and providing multi-model cooperative detection attack with weight optimal distribution. The multi-model cooperative defense method is high in safety and can effectively defense the attack of a deep learning model for the antagonism attack.