Label-independent cross-task adversarial sample generation method

The invention discloses a label-independent cross-task adversarial sample generation method, and the method comprises the following steps: 1, building a replacement model: obtaining a label set of a pure sample set, and obtaining the replacement model of a target model through the training of the pure sample set and the label set; step 2, generating an adversarial sample; inputting the pure sampleset into a substitution model; obtaining feature representation of each sample in the pure sample set in the substitution model; and calculating the adversarial loss of each sample by utilizing the characteristic representation and calculating the standard deviation, updating each sample by utilizing the adversarial loss and calculating the gradient, and repeatedly executing the step 2 on each sample in the pure sample set for multiple times to obtain an adversarial sample set of the pure sample set. Through the label-independent cross-task adversarial sample generation method, the generatedadversarial sample has transferability, high relativity and task independence.