AdvDPDiff: A diffusion model adversarial attacks strategy combining multi-model ensemble and dynamic input transformation☆

Men, DaPeng

doi:10.1016/j.jvcir.2026.104843

AdvDPDiff: A diffusion model adversarial attacks strategy combining multi-model ensemble and dynamic input transformation☆

2026-08-01

0

PRE

AI

DOI：10.1016/j.jvcir.2026.104843

Original

Original request for help

Save

Abstract

En 中文

Diffusion models have recently proven effective at generating high-quality adversarial examples, yet existing approaches struggle with transferability across heterogeneous models and instability in loss-function optimization. To address these challenges, we propose AdvDPDiff, a novel strategy that combines a multi-model ensemble with dynamic input transformations. AdvDPDiff integrates three modules: the Multi-Model Collaborative Perturbation Efficient Ensemble Module (PEEM), the Dynamic Input Diversity Transformation Module (DIDT), and the Loss Normalization Regularization Strategy (LNOS). The PEEM enhances perturbation transferability by optimizing gradients from multiple models. The DIDT enriches input diversity through random geometric transformations, improving robustness. LNOS stabilizes training by balancing loss terms. Extensive experiments demonstrate that AdvDPDiff outperforms existing attack methods in transferability and robustness across diverse models, demonstrating its superior performance in black-box settings.

Keywords:

Diffusion models

Image classification

Adversarial examples

Black-box attacks

Transferability enhancement

AdvDPDiff: A diffusion model adversarial attacks strategy combining multi-model ensemble and dynamic input transformation☆

Abstract

Journal

Organization

Cited Papers