Adversarial sample detection method and universal adversarial attack defense system

• Main Authors: YE JIAQUAN, WU HEFENG, LIN JING, WANG QING
• Format: Patent
• Language: Chinese; English
• Published: 23.02.2021
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; HANDLING RECORD CARRIERS; PHYSICS; PRESENTATION OF DATA; RECOGNITION OF DATA; RECORD CARRIERS

The invention discloses an adversarial sample detection method, and the method comprises the steps: obtaining a training data set for training a deep neural network model, and obtaining a prediction unit A; utilizing an adversarial sample generated based on the training data set to train a deep neural network model through an adversarial training method to obtain a prediction unit B; inputting thetraining data set and the adversarial sample into prediction units A and B for reasoning, respectively extracting feature maps output by the same convolution layer and splicing the feature maps, andtaking the spliced map as a classification training data set; training a deep neural network binary classification model by adopting the classification training data set to obtain an adversarial sample detection module; and respectively inputting input samples to be detected into the prediction units A and B for reasoning, respectively extracting feature maps output by the same convolution layer and splicing the feature ma