Robust Deep Convolutional Neural Network inspired by the Primary Visual Cortex

Abstract

Most of the current advanced object recognition deep convolutional neural networks (DCNNs) are vulnerable to attacks of adversarial perturbations. In comparison, the primate vision system can effectively suppress the inference of adversarial perturbations. Many studies have shown that the fusion of biological vision mechanisms and DCNNs is a promising way to improve model robustness. The primary visual cortex (V1) is a key brain region for visual information processing in the biological brain, containing various simple cell orientation selection receptive fields, which can specifically respond to low-level features. Therefore, we have developed an object classification DCNN model inspired by V1 orientation selection receptive fields. The V1-inspired model introduces V1 orientation selection receptive fields into DCNN through anisotropic Gaussian kernels, which can enrich the receptive fields of DCNN. In the white-box adversarial attack experiments on CIFAR-100 and Mini-ImageNet, the adversarial robustness of our model is 21.74% and 20.01% higher than that of the baseline DCNN, respectively. Compared with the SOAT VOneNet, the adversarial robustness of our model improves by 2.88% and 8.56%, respectively. It is worth pointing out that our method will not increase the parameter quantity of the baseline model, while the extra training cost is very little.