Method for defense of attack of adversarial examples based on convolutional denoising auto-encoder

Abstract

The present invention relates to a method for defense of attack of adversarial examples based on a convolutional denoising auto-encoder. Adversarial image examples x* (output tags identified by an image classification device is y*) are constructed by manual addition of adversarial disturbance on clean image samples x without modification (output tags identified by the image classification device is y), the fraud purpose that y* is not equal to y can be achieved, even though the image classification device classifies two images essentially showing the same meaning to two classes by mistake. Thepresent invention designs an integration defense model connected with a target image classifier based on a convolutional denoising auto-encoder (CDAE), namely input samples are subjected to coding and decoding at the internal portion of a well trained CDAE to remove most of adversarial disturbances in the input samples so as to output denoising samples close to original clean samples, and then are transmitted to the target image classifier so as to improve the classification correction of the target classifier and have an effect for defense of attack of adversarial examples.

Description

technical field [0001] The invention belongs to the technical field of information security, and relates to a method for defending against an adversarial sample attack based on a convolution denoising autoencoder. Background technique [0002] As machine learning technology is widely used in various fields, including identity verification, automatic driving, speech recognition and other fields, its security has also attracted everyone's attention. Nguyen et al. found in 2014 that deep neural networks are easily fooled by adversarial examples. In 2015, Goodfellow et al. showed that any machine learning classifier can be fooled by adversarial examples, not limited to deep learning networks. The attacker slightly modifies the input data source so that the user cannot perceive it, and realizes that the machine learning system accepts the data and makes wrong follow-up operations, that is, in the unmodified clean sample x (image classifier recognition output The adversarial ima...

AI Technical Summary

Problems solved by technology

Compared with the general training process, the addition of adversarial samples means that the training process needs to consume more computing resources and spend more time; and it is difficult to fit both adversarial samples and clean samples during training; in addition, the effect of adversarial training Depending on the representativeness of the adversarial sample set, adding new adversarial samples requires retraining the target model to consolidate the defense, so the efficiency is not good; there is also the problem of lack of good interpretability

Images