To Make Yourself Invisible with Adversarial Semantic Contours
Published in Computer Vision and Image Understanding (CVIU), Volume 230, 2023
This paper proposes to take use of object semantic contours to enhance the sparse attack, i.e., $\ell_0$-constrained attack, on modern object detectors. We design a prior over the object contour for pixel selection and defines Adversarial Semantic Contours (ASC) as a Maximum A Posteriori (MAP) estimate under a Bayesian formulation of $\ell_0$-constrained sparse attack. Comprehensive experiments are conducted to verify the attack performance of ASC with 9 modern detectors in 3 datasets.
