Method for resisting attack deep neural network based on frequency band

Abstract

The invention discloses a method for resisting an attack deep neural network based on a frequency band, and the method comprises the steps: firstly obtaining a frequency band space which comprises theanti-interference of frequency band pairs in N directions, wherein the anti-interference of frequency band pairs in each direction comprises the anti-interference distribution of frequency band pairswith M wavelengths, the anti-interference distribution of each frequency band pair is a waveform containing alternating activation and inhibition, and gradient signals are filled between wave crestsand wave troughs of each frequency band pair; respectively adding the frequency band pair anti-interference distribution and the original image to obtain an anti-attack sample, inputting the anti-attack sample into the attacked deep neural network, taking the anti-attack sample corresponding to the minimum value of the prediction result score as an optimal anti-attack sample, and attacking the attacked deep neural network. According to the method, the waveform containing alternate activation and suppression is adopted as anti-interference distribution, so the sensitivity of the network to an anti-attack sample is effectively improved, the anti-attack sample can be easily measured by the deep neural network, the original signals are confused, and the deep neural network is enabled to predict an error result.

Description

technical field [0001] The invention belongs to the technical field of computer vision, and in particular relates to a frequency-band-based anti-attack deep neural network method. Background technique [0002] With the development of deep neural networks, recognition tasks such as classification, segmentation, and skeleton detection in computer vision have been greatly improved. Deep neural networks have shown excellent performance in visual recognition tasks, however, when they generalize to other distributions of data, the performance degrades significantly. The generalization of the deep neural network is very critical in its application to practical situations. If the generalization of the deep neural network is poor, it will lead to major errors in actual use and cause serious losses. Therefore, before the deep neural network is applied, its generalization ability needs to be studied clearly. [0003] Adversarial attack is an important method to detect the generalizat...

AI Technical Summary

Problems solved by technology

[0007] Aiming at the technical problems existing in the prior art, the present invention provides a frequency band-based anti-attack deep neural network method to solve the technical problem that the anti-jamming law generated by the existing anti-attack method is still difficult to explain

Images