Adversarial attack via dual-stage network erosion

Deep neural networks are vulnerable to adversarial examples, which can fool deep models by adding subtle perturbations. Although existing attacks have achieved promising results, it still leaves a long way go for generating transferable examples under the black-box setting. To this end, paper proposes improve transferability of and applies dual-stage feature-level perturbations an model implicitly create set diverse models. Then these fused longitudinal ensemble during iterations. The proposed method is termed Dual-Stage Network Erosion (DSNE). We conduct comprehensive experiments both on non-residual residual networks, obtain more with computational cost similar state-of-the-art method. In particular, be significantly improved biasing block information skip connections. Our work provides new insights into architectural vulnerability presents challenges robustness networks.