Train a Highly Robust Model with a Data-enhanced Wavelet Residual Network

Abstract

Once the classification neural network is attacked by adversarial attack, the classification accuracy will drop significantly. Therefore, the combination of wavelet transform and ResNet18 network model, and then linear interpolation adversarial training is used to train the defense model to improve its classification accuracy. Before training the network, the samples are linearly interpolated to obtain the interpolated samples and interpolation labels, and after the image features are convolved and normalized layers, the wavelet transform is applied to downsample to help filter out the high-frequency information, and finally the interpolation loss is obtained to update the network parameters. The trained model is tested using multiple adversarial attack methods on the CIFAR-10 and SVHN datasets, and the experimental results show that the proposed method can show excellent defense performance under strong attacks, and improve robustness and training time.