Quantum-Inspired Frequency Attenuation for Enhanced Targeted Fabrication Attacks in Object Detection

Compared to untargeted attacks, targeted attacks for object detection are much more challenging. Most existing targeted attack methods focus on the object-mislabeling targeted attack, while few have explored the object-fabrication targeted attack, which can mislead detectors into fabricating extra specified objects. However, optimizing such attacks is hindered by local optima due to redundant image details. In this paper, we propose a novel Quantum-Inspired Frequency Attenuation (QIFA) method, drawing inspiration from the barrier-penetration view of quantum tunneling effect. QIFA stochastically eliminates redundant high-frequency information, enhancing the optimization process and improving attack performance. Additionally, we implement a dual attack strategy on feature and output layers to leverage complementary model information, further boosting effectiveness. Extensive experiments on the MS COCO dataset with diverse object detectors demonstrate that our method significantly outperforms state-of-the-art targeted attack approaches in both white-box and black-box attack settings.