Robustness image classification method based on multi-model adversarial distillation

Abstract

The invention provides a robustness image classification method based on multi-model adversarial distillation, and the method comprises the following steps: S1, obtaining a data set, and carrying out the pre-training of a complex model, and obtaining a model T2; s2, according to the training data set, generating corresponding adversarial sample data through an adversarial sample generation method; s3, inputting an adversarial sample into the complex model for adversarial training to obtain a model T2; and S4, selecting a lightweight model which has the same task as the complex model as a student model S, and performing distillation training on the student model through a multi-model knowledge distillation framework. The robust image classification method based on multi-model adversarial distillation is realized through adversarial training and knowledge distillation methods, students can fully learn characteristics of different teacher models through knowledge distillation, and precision is considered while adversarial robustness is improved.

Description

technical field [0001] The invention relates to a robust image classification method, in particular to an image classification method based on a deep neural network, and belongs to the fields of deep learning and artificial intelligence. Background technique [0002] In recent years, with the rapid development of deep learning technology, multi-layer neural network structures emerge one after another, and deep neural networks have achieved unprecedented results in image classification tasks. Neural network model design should achieve a good trade-off between model performance and model complexity. However, in practice, it is difficult for researchers to determine the right balance point, so they prefer to choose neural network models that are over-parameterized, have sufficient expressive ability, and are easy to optimize. Although the neural network can have better expressive ability with increasing depth and more complex structure, it also faces more resource consumption....

AI Technical Summary

Problems solved by technology

[0006] The present invention overcomes the defects of complex model structure and low generalization ability of the classification model in the prior art, and provides a robust image classification method based on multi-model anti-distillation

Images