AI 기반 실시간 해안 감시체계 우회를 위한 적대적 예제 공격 위험성 연구
A Research on the Risks of Adversarial Examples Attacks to AI-Based Real-Time Coastal Surveillance Systems
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
인공지능(Artificial Intelligence: AI) 기술이 국방 분야에서도 활발히 적용되고있다. 특히, 딥러닝(Deep Learning: DL) 기반의 객체 인식 모델을 활용해2029년부터 무인 경계 감시와 피아식별 자동화가 이루어질 예정이다. 이러한 DL 기반의 객체 인식 기술은 기존의 감시체계에 비해 효율성과 정확성을 크게향상시킬 잠재력을 가지고 있는 반면에 적대적 공격(Adversarial Attack)에취약하다는 문제도 함께 지적되고 있다. 특히, 적대적 예제 공격(Adversarial Example Attack)을 통해 객체 탐지 모델의 피아식별 기능을 교란할위험성이 크다. 본 연구에서는 가까운 미래에 우리가 운용하게 될 AI 기반의해안감시시스템을 대상으로 적대적 예제 공격 기법을 통해 침투를 시도하려는군함의 객체 탐지 모델의 성능 저하를 유도하고, 탐지 정확도에 미치는 영향을실험을 통해 분석함으로써, AI 객체 탐지 모델이 우리에게 제공하는 잠재적 이점을극대화하기 위해서는 적대적 예제 공격에 대한 모델의 취약성을 면밀히 분석하고, 이를 보완할 실질적인 대응 전략을 마련하는 것이 필수적임을 강조하고자 한다.
Abstract
Artificial Intelligence (AI) technologies are increasingly being applied in the defense sector. In particular, object recognition models based on deep learning (DL) are expected to facilitate automated unmanned surveillance and friend-or-foe identification starting in 2029. However, DL-based object recognition technology is known to be vulnerable to adversarial attacks, especially adversarial example attacks, which pose significant risks of disrupting the friend-or-foe identification capabilities of object detection models. This study focuses on analyzing the potential risks posed by adversarial example attacks on AI-based coastal surveillance systems that may be deployed in the near future. Specifically, the research examines how such attacks, aimed at warships attempting to infiltrate, could degrade the performance of object detection models and impact detection accuracy. By conducting experiments using transfer-learned models on real-world datasets, this study aims to raise awareness about the risks of adversarial example attacks on AI object detection models and emphasize the critical need for effective defense measures against such threats.
참고문헌 (14)
- Jung Hyun Yoon, "Major issues in the introduction of artificial intelligence technology in the defense sector and measures to enhance utilization", STEPI Insight, 279, pp.1~55 (2021)
- Joseph Redmon, "You Only Look Once : Unified, Real-Time Object Detection", Proceedings of the IEEE conference on computer vision and pattern recognition, pp.779~788 (2016)
- Ian Goodfellow, "Explaining and harnessing adversarial examples", ICLR 2015 conference (2014)
- Eykholt, Kevin, "Robust physical-world attacks on deep learning visual classification", Proceedings of the IEEE conference on CVPR (2018)
- S Povolny, "Model Hacking ADAS to Pave Safer Roads for Autonomous Vehicles", McAfee Advanced Threat Research (2020)
- "Ultralytics YOLOv11"[URL]
- Sandip kundu, "Security and Privacy of Machine Learning Algorithms", ISQED, ~ (2019)
- "Twitter"[URL]
- Mozaffari-Kermani, "Systematic poisoning attacks on and defenses for machine learning in healthcare", IEEE journal of biomedical and health informatics, 19(6), pp.1893~1905 (2015)
- Kevin Eykholt, "Robust Physical-World Attacks on Deep Learning Visual Classification", Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp.1625~1634 (2018)
- Qing Guo, "Learning to Adversarially Blur Visual Object Tracking", IEEE, 2021, 10839~10848
- Eykholt, Kevin Ivan Evimov, "Physical Adversarial Examples for Object Detectors", 12th USENIX(WOOT 18) (2018)
- "Roboflow Data"[URL]