A Frequency Band-Based Adversarial Attack on Deep Neural Networks

Abstract

The invention discloses a method for counterattacking a deep neural network based on a frequency band. First, the frequency band space is obtained, which includes the frequency band countermeasures in N directions, and the frequency band countermeasures in each direction include the frequency band countermeasures of M wavelengths. Interference distribution; the anti-interference distribution of each frequency band is a waveform containing alternating activation and suppression, and the gradient signal between the peaks and the troughs is filled with gradient signals; the frequency-band anti-interference distribution is added to the original image to obtain the adversarial attack sample, which is input to the target. In the attacking deep neural network, the adversarial attack sample corresponding to the minimum value of the predicted result score is taken as the optimal adversarial sample, and the attacked deep neural network is attacked; the present invention adopts the waveform including the alternating activation and suppression as the adversarial interference distribution, which effectively The sensitivity of the network to adversarial attack samples is improved, and the adversarial attack samples can be easily detected by the deep neural network, confuse the original signal, and make the deep neural network predict wrong results.

Description

technical field [0001] The invention belongs to the technical field of computer vision, and in particular relates to a frequency band-based confrontation attacking 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 perform well in visual recognition tasks, however, when generalized to other distributions of data, performance degrades dramatically. The generalization of deep neural network is very critical in its application to practical situations. If the generalization of deep neural network is poor, it will lead to major errors and serious losses in actual use. Therefore, before the application of deep neural network, its generalization ability needs to be clearly studied. [0003] Adversarial attack is an important method to detect the generalization and security of deep ne...

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