미디어 콘텐츠 보안을 위한 양방향 객체 추적 및 최적 식별 모델 조합 기반의 강건한 동적 워터마킹 연구

초록

This paper proposes a robust dynamic watermarking pipeline based on a bidirectional object tracking algorithm to overcome thelimitations of conventional watermarking in dynamic media environments. By integrating YOLOv8 and CLIP, the system achievessuperior identification robustness in the video security domain compared to general-purpose zero-shot models like OWL-ViT2. Specifically, we designed a ‘Reference-based Bidirectional Temporal Propagation’ logic using SAM2, ensuring precise maintenanceof watermark masks even amidst complex movements and occlusions. Furthermore, technical rigor was enhanced through averification framework based on Hamming distance to mitigate bit errors caused by video manipulation. Experimental results usingthe OTB benchmark demonstrated that the proposed method achieved a 2.6-fold increase in tracking IoU compared to the baselineand maintained a detection rate exceeding 96% under various distortion attacks. This study provides a practical technical foundationfor real-time broadcasting security by identifying the optimal combination of state-of-the-art vision models.

키워드