.png){kind=icon}
.png){kind=icon}
Return
MWDP: Multi-view wavelet-guided diffusion purifier for robust pattern recognition
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2026.133041.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Diffusion purifiers (DPs) have emerged as promising pre-processing modules for enhancing the adversarial robustness of machine learning models. While recent advances in frequency-guided DPs leverage wavelet methods to boost purification, these models suffer from latent semantic inconsistency and limited frequency awareness. To address these limitations, we propose the Multi-view Wavelet-guided Diffusion Purifier (MWDP), which integrates two key strategies: (i) multi-view fusion leveraging depth and near-infrared features as structural priors to guide semantic consistency, and (ii) soft-mask learning to capture frequency-band sensitivity. MWDP is plug-and-play, requiring no modifications to target models, and achieves a favorable trade-off between accuracy and time-efficiency even under strong adaptive attacks. Experimental results show that MWDP surpasses state-of-the-art baselines such as DiffPure and IWMF-Diff, demonstrating superior cross-model generalization and practical potential.
Keywords:
Diffusion purifier
adversarial robustness
wavelet-guided
multi-view fusion
soft-mask learning
Journal
IF:
6.5
Papers:
2.5W
Citations:
6.5W