Black box adversarial sample generation method based on microbial genetic algorithm

Abstract

The invention relates to a black box adversarial sample generation method based on a microbial genetic algorithm, and belongs to the technical field of artificial intelligence safety. The method mainly solves the problem of excessive number of times of query required for successfully attacking a neural network model to generate an adversarial sample under the condition of a black box, combines twotypical methods in black box attacks, namely migration attacks and output-based attacks, and solves the discretization problem by using a simple microbial genetic algorithm.

Description

technical field [0001] The invention belongs to the technical field of artificial intelligence security, and in particular relates to a method for generating a black-box confrontation sample based on a microbial genetic algorithm. Background technique [0002] Deep learning has made significant progress in image classification, speech recognition, machine translation, face recognition, and object detection. However, deep learning is easily fooled by adversarial examples. Adversarial examples are generated by adding some interference that humans cannot detect in the normal input. For example, we input a picture of a cat. Under normal circumstances, the image classification model will classify it into the category of cat, but if we are in Add some carefully crafted noise to this image, and the image classification model is likely to classify it into a different class. Since the discovery of adversarial examples, serious security problems have emerged in many critical domains...

AI Technical Summary

Problems solved by technology

[0005] The present invention mainly solves the problem of too many queries required for successfully attacking the neural network model to generate adversarial samples in the case of a black box, and proposes a black-box adversarial sample generation method based on microbial genetic algorithm

Images