TransPatch: A Transformer-Based Generator for Accelerating Transferable Patch Generation in Adversarial Attacks Against Object Detection Models

Abstract

Patch-based adversarial attack shows the possibility to black-box physical attacks on state-of-the-art object detection models through hiding the occurrence of the objects, which causes a high risk in automated security system relying on such model. However, most prior works mainly focus on the attack performance but rarely pay attention to the training speed due to pixel updating and non-smoothing loss function in the training process. To overcome this limitation, we propose a simple but novel training pipeline called TransPatch , a transformer-based generator with new loss function, to accelerate the training process. To address the issue of unstable training problem of previous methods, we also compare and visualize the landscape of various loss functions. We conduct comprehensive experiments on two pedestrian and one stop sign datasets on three object detection models, i.e., YOLOv4, DETR and SSD to compare the training speed and patch performance in such adversarial attacks. From our experiments, our method outperforms previous methods within the first few epochs, and achieves absolute $20\% \sim 30 \%$ 20 % ∼ 30 % improvements in attack success rate (ASR) using $10\%$ 10 % of the training time. We hope our approach can motivate future research on using generator in physical adversarial attack generation on other tasks and models.