AttPD: Adversarial Patch Integrating Attention Mechanism and Dropout

Adversarial attacks aim to deceive deep neural networks by introducing deliberately crafted perturbations, causing them to produce incorrect outputs and posing a significant threat to the security of neural networks. Existing black-box adversarial attack methods are primarily categorized into query-based and transfer-based approaches. However, query-based methods face high query costs, while transfer-based methods often rely on model ensemble techniques to enhance transferability, leading to high computational overhead. To address these challenges, this paper proposes an adversarial patch generation method based on an attention mechanism, called AttPD. By designing an attention loss function to perturb the attention distribution of target models, AttPD significantly improves the cross-model transferability of adversarial patches without substantially increasing computational costs. Furthermore, inspired by the Dropout mechanism in neural network training, this paper introduces a novel Patch Dropout method. By randomly masking parts of the adversarial patch, this method simulates potential wear or damage to the patch in real-world scenarios, enhancing the robustness of adversarial patches. Additionally, we construct a dataset named AttData, consisting of 23 target classes, to comprehensively validate the effectiveness of AttPD. Detailed experiments conducted on multiple object detection models using AttData demonstrate that AttPD achieves strong attack performance in both white-box and black-box scenarios, exhibiting high transferability and practicality.