Confrontation and defense method based on hierarchical classification

The invention discloses a confrontation and defense method based on hierarchical classification, and the method comprises the steps: building a two-stage hierarchical classification model: dividing all samples into N classes in the first stage, and building M classification tasks for the samples predicted for each class in the first stage, namely N M-classification sub-classification tasks, in the second stage; the classifiers comprise a super-class classifier and a subclass classifier, the super-class classifier outputs through a Softmax layer to obtain a super-class probability, and the super-class probability is multiplied by each subclass probability value to which the super-class belongs, so that the prediction of the subclass is standardized to the subclass to which the subclass belongs, and the layering is realized; according to the method, adversarial information is not introduced artificially, and adversarial samples do not need to be generated by spending a large amount of computing resources and ti