Deep learning sample-level adversarial attack defense method and device based on neuron activation pattern

Abstract

The invention discloses a deep learning sample-level confrontation attack defense method based on a neuron activation pattern and a device thereof, comprising the following steps: constructing a deep learning model for target recognition, the deep learning model including a convolution layer, a pooling layer and Classification layer; extract the feature map from the network layer of the deep learning model to construct the neuron activation pattern, the neuron activation pattern and the cross entropy function form the loss function; when using normal image samples to train the deep learning model, the constructed loss function is used to optimize The parameters of the deep learning model; according to the gradient of the loss function, the pixel increment is obtained as the disturbance for defending against attacks; when using the deep learning model with parameter optimization for target recognition, the image to be recognized is added with disturbance and input to the deep learning model, after calculation Get the target recognition result. This method can effectively defend against various adversarial attacks without affecting the correct rate of normal samples.

Description

technical field [0001] The invention belongs to the field of information security, and in particular relates to a deep learning sample-level confrontation attack defense method and a device thereof based on neuron activation patterns. Background technique [0002] Deep learning defines a new data-driven programming paradigm, which uses a set of training data to construct the internal system logic of the neuron network, and learns the internal laws and representation levels of the data from the sample data, which can obtain better than general algorithms. More accurate classification results with strong feature expression ability. Due to the ability of deep learning to match or even exceed human performance, it has been widely used in many safety-critical scenarios. In principle, deep learning models need strong robustness and security in order to well resist any potential risks. [0003] However, deep learning itself lacks interpretability, which means it is vulnerable to ...

AI Technical Summary

Problems solved by technology

These defense methods, to a certain extent, change the feature extraction capabilities of the model for samples, so that the model can re-identify the correct perceptual characteristics of the samples, and the cost will be very huge and not conducive to the application in the deployment stage.

In addition, deep learning models may have a common perturbation that causes classification errors by generating perturbations that cross all class classification boundaries

Traditional defense methods are difficult to effectively defend against such attacks

Images