A Universal Adversarial Attack on CNN-SAR Image Classification by Feature Dictionary Modeling

Abstract

Synthetic aperture radar (SAR) image classification with deep learning methods has achieved high accuracy on a variety of scenes. Despite the excellent performance of new methods, the phenomenon that small perturbations in data might lead to a sharp change in the result, raises attention to these black architectures. Increasing number of adversarial attacks on convolutional neural network (CNN) have been proposed, while these methods construct their adversarial examples with the aid of corresponding classifiers. Such condition cannot be realized in actual confrontation. Therefore, we introduce a universal adversarial attack on CNN-SAR image classification. In essence, this method focuses on distinguishing target distribution by feature dictionary modeling, excluding prior knowledge of any classifier. Experiments on simulated data of plane models indicate that this proposed method works well at various typical CNNs.