Towards Feature Space Adversarial Attack by Style Perturbation

We propose a new adversarial attack to Deep Neural Networks for image classification. Different from most existing attacks that directly perturb input pixels, our focuses on perturbing abstract features, more specifically, features denote styles, including interpretable styles such as vivid colors and sharp outlines, uninterpretable ones. It induces model misclassification by injecting imperceptible style changes through an optimization procedure. show can generate samples are natural-looking than the state-of-the-art unbounded attacks. The experiment also supports pixel-space detection defense techniques hardly ensure robustness in style-related feature space.