Active dropblock: Method to enhance deep model accuracy and robustness

Abstract In this study, we investigated a means to improve the robustness of deep network training on visual recognition tasks without sacrificing accuracy. The contribution of this work can reduce the dependence on model decay to gain a strong defense against malicious attacks, especially from adversarial samples. There are two major challenges in this study. First, the model defense capability should be strong and improved over the training stage. The other is that the degrading of the model performance must be minimized to ensure visual recognition performance. To tackle these challenges, we propose active dropblock (ActDB) by incorporating active learning into a dropblock. Dropblock effectively perturbs the feature maps, thus enhancing the invulnerability of gradient-based adversarial attacks. In addition, it selects an optimal perturbation solution to minimize the objective loss function, thereby reducing the model degradation. The proposed organic integration successfully solved the model robustness and accuracy simultaneously. We validated our approach using extensive experiments on various datasets. The results showed significant gains compared to state-of-the-art methods.