Research On Image Restoration And Adversarial Example Defense Algorithm Based On Convolutional Neural Network

As one of the objective reflection media of vision,image plays an essential role in human visual perception and has extensive applications in computer vision,multimedia,medicine,and other fields.High-quality images indicate rich texture and color information,which not only provide intense visual enjoyment but also offer powerful data support for computer vision applications.However,images are often subject to various factors,such as occlusion and noise,which can result in missing or damaged information,thus reducing image quality and usability.Moreover,adversarial noise in images can pose a significant threat to the application of deep learning techniques in real-life scenarios.Therefore,restoring missing information in images and removing adversarial noise have become important and challenging issues in image processing.In this paper,we address two significant problems in image processing: the completion of missing random pixel information in natural images and the removal of global adversarial noise generated by white-box attacks in natural images.The main research achievements and innovations of this paper are as follows:(1)This paper proposes an image restoration method that combines traditional low-rank matrix completion with convolutional neural networks(CNNs).To address the problem of randomly missing pixel information in natural images,the image restoration problem is first transformed into a matrix completion problem.For the traditional low-rank matrix completion problem,a mathematical optimization model is designed based on matrix decomposition,and the alternating direction method of multipliers(ADMM)is used to iteratively update the variables.Then,we construct a fixed-depth CNN based on the update formula to simulate the iterative updates of the variables.Finally,we learn the underlying relationship between the observed data and recover the complete image information through end-to-end trainable method.Experimental comparisons were conducted on the commonly used Set12 dataset for digital image processing.The results demonstrate that the proposed method outperforms current mainstream low-rank matrix recovery methods,especially for cases with higher missing rates.For example,when the image missing rate is as high as 50%,the proposed method can improve image quality by up to 4.23 d B compared to existing methods for single-image restoration.(2)This paper proposes a denoising defense model for removing adversarial perturbations based on the residual learning strategy and a multi-level representation of images.To address the problem of removing global adversarial noise from images,we introduce a convolutional neural network model based on residual learning to learn perturbation information in adversarial examples.First,the entire network employs a single residual unit to extract the perturbation information.Then,to avoid damaging the intrinsic structure of the adversarial perturbations,single padding operation is used in the network instead of the padding operation in each convolutional layer.Finally,a multi-level representation of the image is utilized to guide the denoising model’s training.Specifically,the reconstruction loss and perceptual loss of the image are jointly used as a training loss function for the denoising model.Unlike pre-processing defense measures that directly reconstruct clean images using neural networks to defend against adversarial examples,this model indirectly extracts the adversarial perturbations by adopting a residual learning strategy,which significantly reduces the difficulty of network training.The joint guidance of the reconstruction and perceptual losses effectively learns the adversarial noise from images without causing too much damage to the original image information,thereby greatly improving the classification accuracy of adversarial examples without sacrificing the classification accuracy of clean images.Extensive experiments are conducted on the CIFAR10,SVHN,and Tiny-Image Net datasets.First,the proposed model is combined with the image restoration method presented in this paper for joint defense,and the results show that joint defense achieves better defense performance.Secondly,based on the residual stacking effect,the proposed model is used to denoise adversarial examples multiple times,and the experimental results show that multiple denoising can further reduce the impact of adversarial perturbations without significantly reducing the classification accuracy of clean images,and even improving it.Finally,compared with other advanced input transformation-based defense methods,the proposed defense method performs better.