Industrial internet anti-attack system and method based on neural architecture evolution

The invention discloses an industrial internet anti-attack system and method based on neural architecture evolution, and the method employs the data to train a replacement model under the condition of training data or data distribution in an industrial internet intrusion detection system based on a deep learning model (i.e., a target model). Designing a Jacobian saliency map attack method to add disturbance to specific features of attack samples in industrial internet data so as to generate confrontation samples capable of avoiding replacement model detection to the greatest extent; wherein the replacement model is formed by combining at most four neural network basic modules, a combination mode is coded, a difference value of classification accuracy of a target model for generating an adversarial sample for an attack sample and the replacement model is taken as individual fitness, and an optimal replacement model with a maximum classification accuracy decline value is obtained through population evolution op