Fool the COOL - On the Robustness of Deep Learning SAR ATR Systems

Over the last years, deep learning automatic target recognition systems have become very popular for synthetic aperture radar images. These systems achieve very high classification rates with common datasets, like the Moving and Stationary Target Acquisition and Recognition (MSTAR) data. A point that is normally not considered is the robustness of these systems, which typically use a softmax layer without rejection class for classification. It has been reported in the past that small variations in the training and test data of deep neural networks might lead to a change in the result. To avoid this situation, several methods to increase the robustness are presented in this paper. These methods vary from simple, like training with noisy samples, to changes in the network structure, particularly the Competitive Overcomplete Output Layer (COOL) is proposed. The COOL gives an output value that also represents a confidence, but with a larger variation than the softmax output. To evaluate the robustness, the DeepFool algorithm is used to creates adversarial examples, i.e. images that look similar to the original data, but cause a wrong classification result. This algorithm is applied to the known training data of the different networks.