Self-Attention 기반 GAN 워터마킹의 비가시성과 강인성 개선 연구
Research on Enhancing the Robustness and Imperceptibility of GAN Watermarking Based on Self-Attention
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
디지털 이미지의 저작권 보호를 위한 워터마킹 기술은 이미지의 시각적 품질을 유지하면서도 다양한 공격환경에서도 워터마크를 안정적으로 복원할 수 있어야 한다. 본 연구에서는 Self-Attention 메커니즘을 도입한GAN 기반 워터마킹 모델을 제안한다. 제안된 모델은 Self-Attention을 활용하여 이미지 내 전역적(Global) 관계와 중요한 특징을 학습함으로써 워터마크 삽입 이미지의 비가시성과 강인성을 동시에 개선하였다. 실험결과, JPEG 압축, 가우시안 노이즈, 블러, 크롭, 밝기 및 대비 조정 등 다양한 공격 시나리오에서 Bit Accuracy는 평균적으로 약 7.94% 향상되었으며, PSNR(Peak Signal-to-Noise Ratio)은 약 1.42dB, SSIM(Structural Similarity Index Measure)은 약 0.06 증가하여 비가시성과 강인성 모두에서 기존 모델 대비우수한 성능을 입증하였다.
Abstract
Digital image watermarking technology for copyright protection must maintain the visual quality of images while ensuring robust recovery of watermarks under various attack scenarios. This study proposes a GAN-based watermarking model incorporating a Self-Attention mechanism. The proposed model leverages Self-Attention to learn global relationships and critical features within images, simultaneously enhancing the imperceptibility and robustness of watermarked images. Experimental results demonstrate that, under various attack scenarios including JPEG compression, Gaussian noise, blur, cropping, brightness, and contrast adjustments, the Bit Accuracy improved by an average of approximately 7.94%, while the Peak Signal-to-Noise Ratio (PSNR) increased by approximately 1.42dB and the Structural Similarity Index Measure (SSIM) improved by approximately 0.06, showcasing superior performance in both imperceptibility and robustness compared to existing models.
참고문헌 (9)
- "Copyright Protection and Accountability of Generative AI : Atack, Watermarking and Atribution" (2023)[URL]
- Ray, A., "Recent trends in image watermarking techniques for copyright protection : a survey", Int J Multimed Info Retr, 9, pp.249~270 (2020)[DOI]
- Himanshu Kumar Singh, "Comprehensive review of watermarking techniques in deep-learning environments", Journal of Electronic Imaging, 32(03), pp.1~23 (2023)[DOI]
- "Self-Attention Generative Adversarial Networks" (2018)[URL]
- K. Hao, "Robust image watermarking based on generative adversarial network", China Communications, 17(11), pp.131~140 (2020)[DOI]
- "HiDDeN: Hiding Data With Deep Networks" (2018)[URL]
- "Self-Attention: A Better Building Block for Sentiment Analysis Neural Network Classifiers" (2018)[URL]
- A. Vaswani, "Attention is All You Need", Advances in Neural Information Processing Systems (NIPS), 30 (2017)
- "Supervised GAN Watermarking for Intellectual Property Protection" (2022)[URL]