객체 탐지기에 대한 적대적 패치 공격에 대응하는 경쟁 학습 기반의 범용적인 잡음 주입 기법
Universal Noise Injection Method base on Competitive Learning to Defeat Adversarial Patch Attacks on Object Detector
Recently, many researches have been conducted on deep learning technology for object recognition in the fields of self-driving cars and smart factories. Nevertheless, these object recognition systems are still vulnerable to malicious adversarial attacks. The patch attack, one of the recently prominent adversarial attacks, is an attack that causes malfunction of a deep learning model by injecting noise to the original image in the form of a patch. In this paper, we confirmed that when an adversarial patch attack is performed targeting object detectors YOLO(You Only Look Once), Retinanet, and Faster R-CNN, the object detection performance is significantly reduced. In addition, we proposed an universal noise injection technique based on competitive learning to defeat these adversarial patch attacks. As a result of applying the proposed technique to the object detection model, YOLO, Retinanet, and Faster R-CNN show higher detection performance than the existing countermeasures by achieving 98.0%, 94.6%, and 90.8% mAP performance respectively.
한국어
최근 자율 주행 자동차나 스마트 팩토리 분야에서 객체 인식을 위한 딥러닝 기술에 대한 많은 연구가 진행되고 있으나 악의적인 적대적 공격(adversarial attacks)에 취약하다는 사실이 밝혀졌다. 최근 주목받고 있는 적대적 공격 중 패치 공격(patch attack)은 원래 이미지에 잡음을 패치 형식으로 부착함으로써 딥러닝 모델의 오동작을 유발하 는 공격이다. 본 논문에서는 객체 탐지기인 YOLO(You Only Look Once), Retinanet 그리고 Faster R-CNN을 대 상으로 적대적 패치 공격을 수행하면 객체 탐지 능력이 현저히 저하됨을 실험을 통해 확인하였다. 또한, 적대적 패치 공격에 대응하기 위한 경쟁 학습 기반의 범용적인 잡음 주입(universal noise injection) 기법을 제안하였다. 제안 기 법을 공격 대상이 되는 객체 탐지 모델에 적용한 결과, YOLO, Retinanet 그리고 Faster R-CNN에서 각각 98.0%, 94.6%, 90.8%의 mAP를 달성하였으며 기존 대응 방법과 비교하였을 때 우수한 성능을 보인 것을 확인하였다.
목차
요약 Abstract Ⅰ. 서론 Ⅱ. 객체 탐지기에 대한 적대적 패치 공격 2.1 객체 탐지 알고리즘 2.2 적대적 패치 공격 2.3 적대적 패치 공격 대응 방안 Ⅲ. 범용적인 잡음 주입 기법 Ⅳ. 적대적 패치 공격 실험 및 대응 분석 4.1 실험 환경 Ⅴ. 적대적 패치 공격 실험 및 모델 성능 평가 Ⅵ. 결론 REFERENCES
Ever since next generation convergence technology became one of the most important industries in the nation, computing professionals have encountered a growing number of challenges. Along with scholars and colleagues in related fields, they have gathered in avariety of forums and meetings over the last few decades to share their knowledge, experiences and the outcome of their research. These exchanges have led to the founding of the International Next-generation Convergence technology (INCA) on December 1, 2015. INCA was registered as an incorporated association under the Ministry of Information and Communications. The main purpose of the organization is to improve our society by achieving the highest capability possible in next generation convergence technology.
간행물
간행물명
차세대융합기술학회논문지 [The Journal of Next-generation Convergence Technology Association]