Publication
Adversarial examples crafted in black-box scenarios are affected by unrealistic colors or spatial artifacts. To prevent these shortcomings, we propose a novel strategy that generates adversarial images with low detectability and high transferability. The proposed black-box strategy, GLoFool, introduces global and local perturbations iteratively. First, a combination of image enhancement filters is applied globally to the clean image. Then, local color perturbations are generated on segmented image regions. These local perturbations are dynamically increased for each region over the iterations by sampling new colors on an expanding disc around the initial global enhancement. We propose a version of the method optimized for quality, GLoFool-Q, and one for transferability, GLoFool-T. Compared to state-of-the-art attacks that perturb colors, GLoFool-Q generates adversarial images with better color fidelity and perceptual quality. GLoFool-T outperforms all the black-box methods in terms of success rate and robustness, with a performance comparable to the best white-box methods.