Image classification method for optimizing decision boundaries and enhancing robustness of deep neural network

Abstract

The invention relates to an image classification method for optimizing decision boundaries and enhancing robustness of a deep neural network, which is characterized in that adversarial samples are calculated by calculating a decision domain model of a network, and distribution of the decision boundaries corresponding to the neural network is optimized through adversarial training, so that the robustness of the network is improved. The method specifically comprises the steps that parameters of a trained deep neural network are used for calculating and obtaining a corresponding decision domain model, the model and a training sample are used for determining which decision boundaries corresponding to the network are sensitive to disturbance, and data in a training set is used for calculating adversarial samples of samples around the decision boundaries sensitive to disturbance; the adversarial samples are mixed into training samples, adversarial training is carried out on the network, network parameters are optimized so that decision boundaries move in the opposite direction of the sensitive direction, the sensitivity of the decision boundaries to disturbance is reduced, the deep neural network with high robustness is obtained, and the anti-interference capability of the network in a classification task is improved.

Description

technical field [0001] The invention belongs to the field of artificial intelligence, and in particular relates to an image classification method for optimizing decision boundaries and enhancing the robustness of deep neural networks. Background technique [0002] Deep neural networks can be used to perform classification and regression tasks. Today, deep neural networks perform well in a variety of classification tasks. When the deep neural network performs classification tasks, it mainly relies on the decision domains obtained by dividing the sample space by the deep neural network. Each decision domain corresponds to its own category. When a sample point falls into a certain decision domain, the corresponding sample is classified. The network is classified into categories corresponding to the decision domain. But for some special picture samples, adding some small perturbations to them will cause the network to make wrong judgments on the category of the picture. This k...

AI Technical Summary

Problems solved by technology

[0005] Aiming at the problem that when the existing model classifies images, adding small disturbances that are difficult to detect by the naked eye will change the category and cause inaccurate image classification, the present invention proposes a method to optimize the decision boundary to enhance the robustness of the deep neural network. The innovation of the image classification method is that a new set of methods is proposed to calculate the network decision boundary and its topological properties to construct the decision domain model of the current network. According to the analysis of the distribution of the decision boundary in the sample, an appropriate strategy is selected to adjust the decision. Boundary: By using samples close to the decision boundary to calculate and generate adversarial samples, through confrontation training, adjust the network parameters corresponding to the decision boundary that is sensitive to disturbances, that is to say, the optimized neural network is less sensitive to samples with disturbances added , thus improving the robustness of the network

Images