Semantically Adversarial Learnable Filters

We present an adversarial framework to craft perturbations that mislead classifiers by accounting for the image content and semantics of labels. The proposed combines a structure loss semantic in multi-task objective function train fully convolutional neural network. helps generate whose type magnitude are defined target processing filter. considers groups (semantic) labels prevent filtered {from} being classified with label same group. validate our three different filters, namely detail enhancement, log transformation gamma correction filters; evaluate adversarially images against classifiers, ResNet50, ResNet18 AlexNet, pre-trained on ImageNet. show generates high success rate, robustness, transferability unseen classifiers. also discuss subjective evaluations perturbations.