Attack-resistant defense method based on neuron activation rate

Abstract

The invention discloses an anti-attack defense method based on a neuron activation rate. The method comprises the following steps: 1), obtaining an image data set, carrying out the preprocessing, and constructing a training set and a test set; 2) constructing a classifier model of a CNN network structure; 3) constructing a dichotomy discriminator D, and judging the output N of the classifier model about the input book x; 4) constructing an overall loss function according to the classifier model and the loss function of the dichotomy discriminator D, and determining an overall optimization target; carrying out joint training on the overall model by adopting the training set, and when the overall loss function is converged, stopping training and storing the model; and 5) inputting a test set to test the trained classifier model, if the classification accuracy is lower than a threshold value, modifying the lambda of the overall loss function, and training again according to the step 4) until the classification accuracy is higher than the threshold value. The defense method provided by the invention has good applicability and accuracy, and has a good defense effect on attack resistance.

Description

technical field [0001] The invention relates to a defense method against attack, in particular to a defense method against attack based on neuron activation rate. Background technique [0002] With its superior performance, deep learning has been widely used in many fields such as image classification, target detection, biological information, speech recognition, natural language processing, etc., and has achieved unprecedented success in the field of computer vision. But the vulnerability of deep models to misclassification by perturbations invisible to the human eye has drawn researchers' attention. In the process of gradually replacing human beings for autonomous decision-making by deep models, due to their vulnerability to adversarial attacks and other issues, they have posed serious threats to network security, data security, algorithm security, and information security. The vulnerability of deep models to adversarial examples also hinders the further application of de...

AI Technical Summary

Problems solved by technology

[0007] (1) These defensive measures can only target existing adversarial samples, and have no ability to resist unknown adversarial input

[0008] (2) These methods often fail to achieve better results in the face of black-box attacks

[0009] (3) For migratable attacks, these defenses are compromised

Images