Camouflaged Adversarial Patch Attack on Object Detector

Abstract

Adversarial attacks have received great attentions for their capacity to distract state-of-the-art neural networks by modifying objects in physical domain. Patch-based attack especially have got much attention for its optimization effectiveness and feasible adaptation to any objects to attack neural network-based object detectors. However, despite their strong attack performance, generated patches are strongly perceptible for humans, violating the fundamental assumption of adversarial examples. In this paper, we propose a camouflaged adversarial patch optimization method using military camouflage assessment metrics for naturalistic patch attacks. We also investigate camouflaged attack loss functions, applications of various camouflaged patches on army tank images, and validate the proposed approach with extensive experiments attacking Yolov5 detection model. Our methods produce more natural and realistic looking camouflaged patches while achieving competitive performance.

Application of Metamaterial in Vibration Suppression of Rotor System

“…Since the RPN network generates a large number of candidate fields containing the candidate boxes, the next stage of the network will arrange the candidate boxes for the RPN network according to the confidence level, the candidate boxes above the confidence threshold were selected for the next stage of classification and location regression.BPatch presents an attack method for RPN network filtering mechanism by re-ducing the confidence of the high confidence candidate obtained by the RPN layer and make the candidate boxes that end up in the next layer of the network contain little or no foreground targets. Liu et al [41] add a patch to the picture and treat the patch as a GT (Ground Truth) check box.Back propagation causes the network to optimize the patch directly so that the final detector is affected by the patch, resulting in a detection error. In addition, Wang et al [42] proposed a particle swarm optimization target detection black box attack EA which uses a natural optimization algorithm to guide disturbance generation in place, but this approach is time-consuming.…”

Synthetic Image Generation for Military Vehicle Detection

“…가상데이터 단일 이미지를 인공적으로 생성할 때 레이블링 단계에서의 비용이 실물 이미지에 비해 약 1/100 가량 절감되는 것으로 추정 [1,2] 군용물체탐지용 가상데이터를 사용하여 딥러닝 모델 을 학습시켰을 경우 단일물체 기반 물체탐지 정확도 85 %(AP50)와 64 %(AP75)를 얻을 수 있었다 [15] . 가능함을 확인하였다 [28][29][30][31][32] .…”