Enhancing adversarial transferability by resolving gradient conflicts in multi-input transformations

doi:10.1016/j.asoc.2025.114409

Enhancing adversarial transferability by resolving gradient conflicts in multi-input transformations

2025-12-06

0

PRE

AI

DOI：10.1016/j.asoc.2025.114409

Original

Original request for help

Save

Abstract

En 中文

• First cross-model transfer enhancement by resolving gradient conflicts. • Reveal optimization bottlenecks from gradient cancellation in multi-input transformations. • Global momentum accumulation captures transferable attack directions. • Two-phase optimization strengthens dominant gradients while preserving perturbation diversity. • Lightweight strategy enables low-cost cross-model attacks on CNNs and Transformers.

Enhancing adversarial transferability by resolving gradient conflicts in multi-input transformations

Abstract

Journal

Organization