WaveAttn-AdvGAN: Image adversarial example generation method based on discrete wavelet transform and global grouped coordinate attention
Year: 2025
Type: article
Abstract: Deep neural networks (DNNs) excel at image classification tasks, but their vulnerability to adversarial examples raise serious security concerns. Although existing adversarial attack methods demonstrate certain threat levels in black-box scenarios, they still exhibit limitations in frequency-domain feature utilization, fine-grained attention mechanisms, and multi-objective optimization coordination. This work proposes WaveAttn-AdvGAN, a framework that integrates Discrete Wavelet Transform (DWT) with Global Grouped Coordinate Attention (GGCA) through bidirectional collaborative optimization. The GGCA module employs group pooling and dual-dimensional attention mechanisms to precisely localize critical regions and optimize perturbation distribution. Concurrently, DWT decomposes images to identify high-frequency sensitive subbands, where Gaussian noise injection rectify gradient directions to improve transferability. A multi-objective optimization function combining adversarial loss, attention loss, and perturbation constraints achieves balanced attack efficacy, imperceptibility, and transferability. Experimental results demonstrate a 98.94 % white-box Attack Success Rate (ASR) and 2.39 % average improvement in black-box transfer attacks, with Structural Similarity Index Measure (SSIM), Frechet Inception Distance (FID), and Low-Frequency Distortion (LF) metrics confirming the superior visual stealth of the generated adversarial examples.
Source: Alexandria Engineering Journal
Authors: Xiaoyin Yi, Long Chen, Xiaofei Zhu, Ning Yu
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,900
Award: G7795898257
OpenAlex ID: W4416453308