Abstract: As for the low success rate of open-set adversary sample defense, an auto-encoder based on attention mechanism is designed. The non-local attention based on Gaussian mean and dual-attention were combined and embedded in the auto-encoder ResNet network that deepened the convolutional neural network to improve the success rate of open set adversary sample defense. The experimental results show that compared with OSAD, the proposed method get better success rate on open-set adversarial sample. On CIFAR10 dataset and SVHN dataset, under FGSM attack, AUC-ROC is improved by 2.3 and 13.4 percentage points respectively and the accuracy of closed-set recognition is improved by 4.1 and 0.1 percentage points respectively; under PGD attack, AUC-ROC is improved by 1.0 and 2.8 percentage points respectively, and the accuracy of closed-set recognition is improved by 3.9 percentage points on SVHN dataset.