.png){kind=icon}
Return
MPQD-ADM: High-efficiency adversarial patch defense method based on mixed pixel quality differences
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2026.116425.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Object detection plays a vital role in many safety-critical systems. However, easily implemented adversarial patch attacks in the physical world significantly threaten current state-of-the-art object detectors. Therefore, it is crucial to develop robust defense methods against patch attacks. However, there is a significant lack of research in this area. To this end, this paper proposes a novel defense method, MPQD-ADM, that aims to improve the robustness of object detectors against adversarial patch attacks. The core advantage of MPQD-ADM lies in its patch-agnostic design and high defense efficiency: it can detect and block the entire adversarial patch without any pre-training, even when the patch's shape, size, and location are unknown. Specifically, we propose an adversarial region detection method that identifies potential object-detection regions in the input image that may be hidden by adversarial patches. Next, we propose an adversarial source localization method based on mixed-pixel quality difference analysis, combining discrete cosine transform (DCT) frequency-domain analysis with wavelet transform multi-resolution enhancement to rapidly and accurately pinpoint the location of adversarial patches. Finally, the adversarial patches are effectively removed using a patch location mask, ensuring the object detector correctly recognizes the processed image. Experimental results demonstrate that MPQD-ADM significantly enhances detector robustness compared to existing state-of-the-art methods. Compared to the baseline optimal method, mean average precision (mAP) improves by 6.27%, and defense success rate (DSR) increases by 7.73%. Additionally, MPQD-ADM exhibits low computational overhead, making it well-suited for real-time detection tasks.
Keywords:
Object detection
Adversarial patch defense
Patch-agnostic
Attack detection
Journal
K
IF:
7.6
Papers:
1.3W
Citations:
4.5W
Organization
Cited Papers
No cited papers available