Artificial intelligence-based confrontation sample generation method, device, equipment and medium

Abstract

The invention relates to the field of image detection in artificial intelligence, and provides an artificial intelligence-based confrontation sample generation method, which comprises the steps of obtaining a target image and a confrontation sample generation model; inputting the target image into an encoder network for encoding to obtain a hidden variable, and inputting the hidden variable into a generator network for image generation to obtain a candidate adversarial sample; determining an image perception similarity between the target image and the candidate adversarial sample; when the image perception similarity is smaller than or equal to a preset threshold value, inputting the candidate adversarial samples into a preset image classification model for classification to obtain an output probability; according to the output probability and a real classification label of the target image, determining whether the candidate adversarial sample meets a preset adversarial sample condition; and if yes, determining the candidate adversarial sample as a target adversarial sample of the target image. The accuracy of the adversarial sample is improved. The invention further relates to the field of block chains and medical treatment. The adversarial sample generation model can be stored in the block chain.

Description

technical field [0001] The present application relates to the field of image detection, and in particular to an artificial intelligence-based adversarial sample generation method, device, equipment and medium. Background technique [0002] Adversarial examples refer to deliberately adding some subtle disturbances that cannot be detected by humans to the input samples, causing the model to give a wrong output with a high degree of confidence. Usually, people build an adversarial example generation model, so that the adversarial examples generated according to the adversarial example generation model can make the recognition model make a wrong judgment, and the human eye cannot distinguish the sample image as an adversarial example. [0003] At present, the main adversarial example generation model is based on the image perception similarity of the neural network as a distance measure between images, and the generated adversarial examples have good versatility, but the current...

AI Technical Summary

Problems solved by technology

[0003] At present, the main adversarial example generation model is based on the image perception similarity of the neural network as a distance measure between images, and the generated adversarial examples have good versatility, but the current adversarial example generation model does not measure the generated The distance between the sample in the image pixel space and the original image, but only constraining the perturbation norm of the hidden quantity, cannot accurately control the image quality of the adversarial sample

Images