Efficient transfer attacks via enhancing perturbation robustness
Year: 2024
Type: article
Abstract: With the rapid development of deep learning technology, artificial intelligence (AI) has found wide applications in diverse domains such as image classification, text processing, and autonomous driving. However, the increasing prevalence of security issues cannot be ignored. Studies have shown that deep neural network models face security risks due to adversarial sample attacks. These attacks involve adding imperceptible perturbations to deceive the model’s classification results, exposing vulnerabilities in deep learning model applications. While transfer attack methods offer practicality in real-world scenarios, their current performance in black-box attacks is limited. In this study, we propose a method that combines an attention mechanism and a frequency domain transformation to enhance the robustness of adversarial perturbations, thereby improving the performance of transfer attacks in black-box attack scenarios of deep learning models. Specifically, we introduce the CBAM-ResNet50 enhancement model based on attention mechanisms into transfer attacks, enhancing the model’s ability to identify important image regions. By adding perturbations to these attention-concentrated regions, adversary perturbation robustness is improved. Furthermore, we introduce a method for randomly transforming image enhancement in the frequency domain, which increases the diversity and robustness of adversarial perturbation by distributing perturbations across edges and textures. Experimental results demonstrate that our proposed method, considering both human perceptibility and computational cost, achieves a maximum black-box transfer attack success rate of 60.05%, surpassing the 49.65% success rate achieved by the NI-FGSM method across three models. The average success rate of the five methods exceeds an improvement of 6 percentage points in black-box attacks.
Source: Metaverse
DOI: 10.54517/m.v5i2.2764
Authors: Chengzhi Zhong, Jipeng Hu, Mengda Xie, Meie Fang
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4403713848