Adversarial image detection based on spatial and frequency information

doi:10.1016/j.displa.2025.103262

Adversarial image detection based on spatial and frequency information

2025-10-27

0

PRE

AI

DOI：10.1016/j.displa.2025.103262

Original

Original request for help

Save

Abstract

En 中文

• An attention based feature integration (ABFI) module is designed to enhance the feature interpretability and discrimination robustness by effectively capturing local features, high-frequency details, and identifying repetitive textures or patterns in the image. • A three-branch adversarial sample detection method (AID-SFI) with ABFI modules is proposed by extracting the features in frequency, global context and local structure domains. The ABFI modules are used to generate intra-domain information in frequency or global context, and also fuse information across the aforementioned three domains. • Extensive experiments conducted on both digital and physical adversarial images detection demonstrate the effectiveness and robustness of the proposed algorithm.

Keywords:

Adversarial example

Adversarial image

Adversarial detection

Attention based feature integration

Adversarial image detection based on spatial and frequency information

Abstract

Journal

Organization