Adversarial example defense system and method for artificial intelligence classification

Abstract

The present invention relates to an adversarial sample defense system and method for artificial intelligence classification, wherein the system includes a first conventional convolutional neural network, a defense model training module, a first downsampling convolutional neural network connected to a first downsampler; a defense model The data sets in the training module are respectively input to the first conventional convolutional neural network, and are input to the first downsampling convolutional neural network through the first downsampler; the adversarial sample training set acquisition module is input to the second In the conventional convolutional neural network and the second downsampler, the second downsampler outputs the defense model after inputting the fully connected model fuser through the second downsampling convolutional neural network together with the second conventional convolutional neural network; the first The conventional convolutional neural network and the first downsampled convolutional neural network correspond to output parameters to the second conventional convolutional neural network and the second downsampled convolutional neural network, respectively. The invention can effectively improve the accuracy and robustness of the classification of adversarial samples.

Description

technical field [0001] The invention relates to a system and method for data model learning, in particular to an adversarial example defense system and method for artificial intelligence classification. Background technique [0002] In recent years, under the background of big data, machine learning has received extensive attention and has achieved great progress in application development in many different fields. For example: financial and economic fields (quantitative analysis, risk control, etc.), unmanned driving fields (environment perception, human-vehicle interaction, route planning, etc.), medical fields (drug mining, disease analysis, medical imaging, etc.), network security fields, etc. (Filter spam, DNS malicious domain name analysis, etc.). However, while machine learning brings us great convenience, it also exposes great security problems. [0003] In the paper published by Christian Szegedy et al. in ICLR2014, the concept of adversarial samples was first pro...

AI Technical Summary

Problems solved by technology

However, this method will lose a certain amount of clean sample classification accuracy

Images