Image recognition defense method based on trap structure

Abstract

The invention discloses an image recognition defense method based on a trap structure. The method comprises the steps of adding the trap structure for guiding the connection and transmission directionof a malignant image to a hidden layer and / or an output layer of an image recognition network constructed based on a neural network; inhibiting trap structure activity, and performing image recognition task training on the image recognition network added with the trap structure by utilizing a normal image; for different types of defense tasks, setting a suppression trap to be in an activated state, and selecting malignant images corresponding to the defense tasks to train the defense tasks for the image recognition network added with the trap structure; and during application, inputting a to-be-detected image into the image recognition network trained through the image recognition task and the defense task, when the activation state value of the trap structure is larger than a threshold value, the to-be-detected image being a malignant image, deleting the malignant image, and achieving the defense effect on image recognition. And potential safety hazards caused by misclassification ofmalignant images are prevented.

Description

technical field [0001] The invention belongs to the field of image recognition, and in particular relates to an image recognition defense method based on a trap structure. Background technique [0002] Deep learning can learn to calculate the potential connection of a large amount of data, obtain more accurate image recognition results than general algorithms, and has powerful feature learning capabilities and feature expression capabilities. The core of deep learning is to use a neural network with huge parameters for feature extraction. Typical neural networks include convolutional neural network (CNN) and recurrent neural network (RNN). [0003] In the ImageNet competition over the years, the deep learning network structure has been continuously developed. For example, the convolutional neural network has continuously developed from AlexNet and VGGNet to ResNet and InceptionNet; in the target detection task, such as based on R-CNN, FastR-CNN, etc. The target detection mo...

Images