Adversarial images with downscaling
Résumé
Most works on adversarial attacks consider small images whose size already fits the model. This paper explores attacking large images on classifiers with different input sizes. Downscaling is a necessary first step to adapt the size of the image to the model that might reform the adversarial signal. This paper studies the possibility of forging adversarial images through different interpolation methods, the distortion of their adversarial signal, and the transferability over other downscaling methods. This paper finally explores attacking an ensemble model which gathers different resizing interpolations to increase the transferability of the attack against a set downscaling kernels.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|