국내기사
국방 디지털 위장무늬와 시각적 은밀성을 고려한 Diffusion 기반 적대적 패치 설계 및 YOLOv5 탐지 회피에 관한 연구 = Design of diffusion-based adversarial patches considering military digital camouflage and visual stealth, and study on YOLOv5 detection evasion
The advancement of AI-based object detection systems in modern warfare has significantly reduced the effectiveness of traditional camouflage patterns, even those that are difficult to identify with the human eye. In response to these changing battlefield environments, new approaches are required to minimize the risk of enemy identification. This study proposes a method for designing adversarial patches similar to military digital camouflage patterns. Existing adversarial patch research primarily employs artificial geometric patterns that are easily identifiable by human observation in actual military environments, presenting significant limitations. To address this issue, we utilized a diffusion-based patch generation architecture to create natural patches that harmonize with actual military uniforms. Experimental validation targeting YOLOv5 models demonstrated the effectiveness of degrading person detection performance. Results showed that the proposed patches achieved comparable attack success rates to existing methods while proving superior stealth against human visual identification through camouflage performance evaluation metrics. This research provides a practical military camouflage solution that simultaneously satisfies AI detection evasion and human visual concealment requirements.