Adversarial sample defense method based on trap type integrated network

The invention discloses an adversarial sample defense method based on a trap type integrated network, and belongs to the field of deep learning and artificial intelligence safety. The method comprises the following steps that: different trap data sets are selected according to an actual application scene, a basic network model and a training data set; generating a trap network of an augmented output category based on the basic network model,training the trap network, screening the trap network of the augmented output category under the standard of accuracy and model diversity, and finally forming a trap type integrated network; and adversarial samples are generated for the trap type integrated network, and using the screened adversarial samples to perform adversarial training to continuously improve the robustness of the trap type integrated network. The efficient adversarial sample defense method provided by the invention is a novel model expansion method, and meanwhile, the defense range of the model is exp