Digital image confrontation sample defense method based on truncated singular values and pixel interpolation

Abstract

Research on resistance machine learning shows that a deep learning architecture is easily attacked by an adversarial sample. The carefully designed small disturbance can cause wrong classification ofnormal images by the neural network, but cannot cause recognition errors to a human vision system. The invention provides a digital image adversarial sample defense method based on truncated singularvalues and pixel interpolation, and the method is based on pixel matrix analysis and image pixel interpolation processing, so as to achieve the purpose of restoring the normal recognition of an imagein a deep learning model. Under the condition of ensuring that normal pictures are not mistakenly classified by the deep learning model due to the operation, countermeasure attacks can be effectivelydefended. According to the method, matrix analysis and singular value decomposition are effectively utilized, an optimal approximation image highly similar to an adversarial sample is reconstructed byreserving a large singular value, and the reconstructed image can be correctly recognized by a deep learning model after pixel interpolation processing. In addition, the unpredictability of the defense model can be improved through pixel interpolation, and therefore the purpose of defending against black box attacks is achieved.

Description

technical field [0001] The invention belongs to the technical field of computer science, and in particular relates to a digital image defense method against samples based on truncated singular values ​​and pixel interpolation. Background technique [0002] Deep learning is at the heart of the current rise of artificial intelligence. In the field of computer vision, deep learning has become the focus of attention. In a challenging large-scale visual recognition task in 2012, Krizhevsky et al. Excellent performance of the Convolutional Neural Network (CNN) model. Thanks to this seminal work, deep learning owes its popularity to a workhorse in applications ranging from self-driving cars to surveillance and security. Although deep neural networks have shown excellent performance in solving complex problems, and have saved a lot of manpower in many fields, they can also achieve higher budget and processing power than manpower. With the rise of technologies such as facial recog...

AI Technical Summary

Problems solved by technology

[0004] In order to solve the above-mentioned problems, the present invention conducts in-depth research on the adversarial perturbation problem faced by the deep learning model. The digital image adversarial example defense method, which is based on the singular value decomposition of the pixel matrix and the pixel interpolation processing of the reconstructed image, can eliminate the adversarial disturbance after ensuring the adversarial disturbance and make the deep learning model can be correctly classified. The normal picture processed by the above method will not affect its correct recognition

Images