GAN 기반 적대적 패턴 생성을 통한 객체 탐지 모델 회피 기법
A GAN-based Adversarial Pattern Generation Method for Evading Object Detection Model
1
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
AI 기술의 발전으로 현대 전장에서 자율무기체계(AWS)의 활용성이 증가함에따라 이에 대한 효과적인 방어와 대응 전략이 요구되고 있다. AWS에 적용된객체 탐지(Object Detection) 기술은 적대적 예제 기법을 통해 우회할 수있다는 점을 활용하여, 아군 자산에 적대적 패턴을 적용함으로써 객체 탐지모델의 탐지를 회피하고 적 AWS의 공격 표적이 되는 것을 방지할 수 있다. 하지만 적의 객체 탐지 기술이 지속 발전할 것을 고려하면 이에 대응하기위해 더욱 강력한 회피 성능을 갖춘 적대적 패턴을 생성해야 한다. 따라서, 본 논문에서는 적대적 생성 신경망(GAN)을 활용한 적대적 위장패턴 생성기법을 제안한다. 실험 결과, 제안 기법이 YOLOv7 기반의 객체 탐지 모델의성능을 효과적으로 저하시킬 수 있음을 확인하였으며 향후 실제 물리적인환경에서의 검증을 통해 AWS의 공격으로부터 아군 자산을 효과적으로보호할 수 있을 것으로 기대된다.
Abstract
Advances in AI technology are increasing the utility of autonomous weapons s ystems(AWS) on the modern battlefield, requiring effective defense and counte rmeasure strategies. Object detection technologies applied to AWS can be eva ded through adversarial example techniques, which apply adversarial patterns t o friendly assets to evade detection by object detection models and prevent th em from being targeted by adversarial AWS. However, considering that the ad versary's object detection technology will continue to develop, it is necessary to generate adversarial patterns with stronger evasion performance to cope wi th it. Therefore, this paper proposes an adversarial camouflage pattern generat ion technique using adversarial generation neural networks(GANs). Experimenta l results show that the proposed technique can effectively degrade the perfor mance of YOLOv7-based object detection models, and it is expected that it ca n effectively protect friendly assets from AWS attacks through future verificati on in real physical environments.
참고문헌 (18)
- U. S. Depart of Defense, "Autonomy in Weapon Systems", DoD Directive, ~ (2023)
- Joana Costa, "How deep learning sees the world: A survey on adversarial attacks & defenses", IEEE Access (2024)
- Simun Yuk, "A Research on Adversarial Example-based Passive Air Defense Method against Object Detectable AI Drone", Journal of Internet Computing & Services, 24(6), pp.119~125 (2023)
- Jeonghun Kim, "Camouflage d Adversarial Attack on Object Detector", 21st International Conference on Control, Automati on and Systems (ICCAS), pp.613~617 (2021)
- Ian Goodfellow, "Generative adversarial nets", Advances in neural information processing systems, 27 (2014)
- Joseph Redmon, "You only look once: Unified, real-time object detection", Proceedings of the IEEE conference on computer vision and pattern recognition (2016)
- Wei Liu, "Ssd: Single shot multibox detector", Computer Vision–ECCV 2016:14th European Conference, Proceedings, Part I 14, pp.21~37 (2016), Springer International Publishing
- Ross Girshick, "Trevor Darrell, and Jitendra Malik. "Rich feature hierarchies for accurate object detection and semantic segmentation", Proceedings of the IEEE conference on computer vision and pattern recognition (2014)
- Chien-Yao Wang, "YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors", Proceedings of the IEEE, 2023, 7464~7475
- Hiskias Dingeto, "Universal Adversarial Training Using Auxiliary Conditional Generative Model-Based Adversarial Attack Generation", Applied Sciences, 13, pp.15~ (2023)
- Aleksander Madry, "Towards deep learning models resistant to adversarial attacks", Proc. Int. Conf. Learn. Represent, pp.1~23 (2018)
- Nicholas Carlini, "Towards evaluating the robustness of neural networks", Proc. IEEE Symp. Secur. Privacy (SP), pp.39~57 (2017)
- Amira Guesmi, "Physical Adversarial Attacks for Camera-Based Smart Systems: Current Trends, Categorization, Applic ations, Research Challenges, and Future Outlook", IEEE Access, 11, pp.109617~109668 (2023)
- Yu-Chih-Tuan Hu, "Naturalistic Physical Adversarial Patch for Object Detectors", IEEE, 2021, 7828~7837