Efficient and transferable reversible adversarial attacks utilizing YUV color space

Yucheng Fan, Zhaoxia Yin, Jiawei Chen, Wanli Lyu

Abstract: Highlights•A dual-strategy design is proposed to eliminate ineffective perturbations in reversible adversarial attacks. First, to address GRP, several Y-channel attack methods (YFGSM, YI-FGSM, YPGD, YMI-FGSM, etc.) are introduced to generate perturbations solely on the Y channel. Second, to avoid EOP, perturbation information is embedded exclusively into the U and V channels, preserving the intact perturbed Y channel during reversible embedding.•To further enhance transferability, an ensemble attack strategy is employed to generate perturbations on multiple models simultaneously, thereby mitigating the overfitting of adversarial perturbations to specific models and resulting in more transferable RAEs.•Experimental results demonstrate that our method not only achieves error-free recovery of the original image, but also delivers high visual quality while achieving high operational speed. Additionally, the generated RAEs exhibit strong transferability across multiple models.

External IDs:dblp:journals/ijon/FanYCL25