Sparse adversarial attack method based on weighted gradient hash activation thermodynamic diagram

Abstract

The invention provides a sparse adversarial attack method based on a weighted gradient Hash activation thermodynamic diagram, which comprises the following steps of: 1, inputting a query video into a video Hash retrieval model to obtain a query video Hash code; 2, acquiring a target video set and respectively inputting target videos in the target video set into the video hash retrieval model to generate a plurality of target video hash codes; and 3, performing point multiplication on the query video hash code and the plurality of target video hash codes, and constructing a Hamming distance function between the query video hash code and the plurality of target video hash codes. According to the sparse adversarial attack method based on the weighted gradient hash activation thermodynamic diagram, the position and the sensitive area of the sparse adversarial attack are determined by using the accuracy of the sensitivity of the weighted gradient hash activation thermodynamic diagram, the pixel cost of the adversarial attack is reduced, and the accuracy and efficiency of sparse adversarial attacks and the imperceptibility of adversarial samples are improved.

Description

Technical field [0001] The present invention relates to the field of video confrontation attacks, and in particular, to a sparse confrontation attack method based on a weighted gradient hash activation. Background technique [0002] The application of depth neural network on hash researcher has greatly improved hash search efficiency. In recent years, the depth neural network has proven to be very fragile under confrontation attack, so security issues related to deep neural networks have attracted people's attention. Research on attacking attacks has also been further developed, and the depth retrieval system will also bear the risk of deep neural network while enjoying the benefits of deep neural networks. [0003] The current confrontation attack method can be divided into two major categories of intensive confrontation attacks and sparse confrontation attacks, compared to intensive attacks, and sparse attacks to achieve attacks by pixel points in the disturbance part, and the ...

AI Technical Summary

Problems solved by technology

[0005] The present invention provides a sparse adversarial attack method based on weighted gradient hash activation heat map, the purpose of which is to solve the problem that traditional adversarial attack methods will generate many redundant pixels, which are not suitable for real scenes and require more pixel costs , leading to the problem of low imperceptibility of adversarial examples

Images