Research On Robust Deep Convolution Neural Network Inspired By The Primary Visual Cortex

Deep Convolutional Neural Networks(DCNNs)are the core of deep learning,which have shown successful results for vision tasks.However,current advanced DCNN models for object recognition can be easily bewildered by imperceptibly small and explicitly crafted perturbations,and can hardly recognize objects in corrupted images.This phenomenon is called adversarial attack.Although,humans have no trouble with such corrupted images,indicating that human visual systems can effectively suppress the inference of adversarial perturbations.Many studies pointed out that the fusion of biological vision mechanisms and DCNN is a promising way to improve the robustness of the model.The primary visual cortex(V1)is a key brain region for cortical visual information processing and contains a variety of simple cell orientationselective receptive fields.The corresponding simple cells can respond specifically to different types of low-level features.Research in the neuroscience field found that many Gaussian convolution kernels are more consistent with the receptive field characteristics of the visual system.Among them,the elliptical anisotropic Gaussian kernels are more similar with the longstrip receptive fields in V1 and can extract the low-level features of the image with a high signal-to-noise ratio.Thus,with the help of multi-scale anisotropic Gaussian kernels,this thesis introduces simple cell orientation-selective receptive fields into the front layer of the DCNN and proposes a robust DCNN model inspired by the Primary visual cortex.The main works and innovations of this thesis are listed as followed:(1)This thesis develops a hybrid DCNN model inspired by the primary visual cortex,which can significantly improve the adversarial robustness of the model.The model consists of a front-end inspired by orientation-selective receptive fields and a neural network back-end built from standard CNN layers.The front-end contains three layers:a convolution layer with multiple biocredible convolutions,a nonlinear layer with simple cell and complex cell nonlinearity,and a V1 neuronal stochasticity layer with a V1 neural noise generator.The front-end is the advantage of the model and the source of adversarial robustness.The image processing of the front-end approximates the V1.Moreover,the front-end does not increase the parameter of the model and can obtain larger gains of adversarial robustness with small additional training costs.(2)In the adversarial robustness analysis experiments on CIFAR-10,CIFAR-100,Mini-ImageNet,and ImageNet,the performance of the model in this thesis far exceeds that of the baseline model,and outperforms the state-ofthe-art V1-inspired model VOneNet in some experiments.Additionally,the combination with the training-based data augmentation can further improve the adversarial robustness of the model.(3)Ablation studies on Mini-ImageNet validate the contribution of the front-end proposed in this thesis to model robustness.The combination of biocredible convolutions and V1 neural stochasticity improves the adversarial robustness jointly.In the model interpretability analysis,the model in this thesis locates the key features for classification with more accuracy and shows a bias toward edge and line features.