Multi-target adversarial patch generation method and device based on integrated attention mechanism

Abstract

The invention belongs to the technical field of artificial intelligence security, and particularly relates to a multi-target adversarial patch generation method and device based on an integrated attention mechanism, and the method comprises the steps: constructing an image classification data set; constructing a multi-target adversarial patch generation framework based on an integrated attention mechanism; defining a loss function and training a multi-target adversarial patch generation framework; and testing the attack effect of the generated countermeasure patch. According to the method, thekey classification area of the input image is positioned by integrating the attention mechanism, so that the better attack performance and mobility of the countermeasure patch are ensured; the inputof the generator makes full use of the original image information, so that the anti-patching effect generated by the generator is better; the input of the generator is also fused with multi-target category information, so that any specified category of the target model can be attacked, and multi-target category attacks can be realized; and the input of the discriminator is cut to ensure that the discriminator learns more context information and improve the visual effect of the adversarial patch.

Description

technical field [0001] The invention belongs to the technical field of artificial intelligence security, in particular to a method and device for generating a multi-objective confrontation patch based on an integrated attention mechanism. Background technique [0002] Deep neural networks have achieved great success in areas such as image classification, image detection, text processing, and speech recognition. As a transformative technology, while bringing huge social and economic benefits, it has also aroused people's concerns and thinking about the safety of artificial intelligence. Previous studies have shown that by adding carefully designed small perturbations to the original samples, deep neural networks can be misjudged. [0003] The existence of adversarial examples poses a major challenge to the application of artificial intelligence, such as automatic driving, face recognition, etc., which prompts scholars to continuously study the attack and defense algorithms o...

AI Technical Summary

Problems solved by technology

But it cannot generate adversarial patches for specific categories

Images