Sparse adversarial attack via robust attack points selection

Yunjian Zhang

doi:10.1016/j.patcog.2026.113490

Sparse adversarial attack via robust attack points selection

2026-03-21

0

PRE

AI

DOI：10.1016/j.patcog.2026.113490

Original

Original request for help

Save

Abstract

En 中文

• This paper proposes a novel attack towards DNNs processing 360∘ images via a perspective-to-sphere transferring strategy. The perturbations are calculated on the multi-perspective planes, and a heuristic migration method is designed to migrate them onto the sphere. • In the perspective domain, a general planar sparse adversarial attack is proposed based on robust attack points selection, with the advantages of small search space and robustness to transformation-based defenses, providing a new perspective on the robustness of deep models. • To find the robust points for launching attacks, a planar robust attack point extraction framework is introduced, and we mathematically prove the relationship between the attack points and loss gradients.

Keywords:

adversarial attack

360∘ images

robust attack points

deep neural networks

perspective-to-sphere transfer

Sparse adversarial attack via robust attack points selection

Abstract

Journal

Organization