Direction-aggregated Attack for Transferable Adversarial Examples

Deep neural networks are vulnerable to adversarial examples that crafted by imposing imperceptible changes the inputs. However, these most successful in white-box settings where model and its parameters available. Finding transferable other models or developed a black-box setting is significantly more difficult. In this article, we propose Direction-aggregated attacks deliver examples. Our method utilizes aggregated direction during attack process for avoiding generated overfitting model. Extensive experiments on ImageNet show our proposed improves transferability of outperforms state-of-the-art attacks, especially against trained models. The best averaged success rate reaches 94.6% three 94.8% five defense methods. It also reveals current approaches do not prevent attacks.