基于交错卷积的潜在空间水印模型
A Latent Space Watermark Model Based on Interleaved Convolution
杜宝泽 1李丽香1
作者信息
- 1. 北京邮电大学网络空间安全学院，北京 100876
- 折叠
摘要
扩散模型在图像生成领域应用广泛，但也带来了内容版权与溯源难题。然而当前扩散模型水印技术普遍存在鲁棒性与保真度的权衡困境，二者难以同时达到理想效果。因此，本文提出了一种通道交错卷积融合机制，进而设计了一种基于交错卷积的潜在空间水印模型，其中采用了一种基于图像域扰动重建的端到端训练范式。所提机制能够实现秘密信息与原始潜在特征的自适应局部感知深度融合，从而在特征层面增强了嵌入效率与表达能力。实验结果表明，所提模型生成的水印图像峰值信噪比高达42dB，优于当前主流基线方法。同时，在多种常见图像扰动条件下，该模型能够保持稳定的水印解码准确率，即在不牺牲鲁棒性的前提下提升了水印的视觉保真度，有效缓解了"鲁棒性-保真度权衡"问题，从而为生成内容版权保护提供了一种可行的解决方案。
Abstract
Diffusion models are widely applied in the field of image generation but also pose challenges in content copyright and traceability. However, current diffusion model watermarking techniques generally face a trade-off between robustness and fidelity, making it difficult to achieve both ideal effects simultaneously. Therefore, this paper proposes a channel-interleaved convolution fusion mechanism and designs a latent space watermarking model based on interleaved convolution, employing an end-to-end training paradigm with image domain perturbation reconstruction. The proposed mechanism enables adaptive local perception and deep integration of secret information with original latent features, thereby enhancing embedding efficiency and expressive capability at the feature level. Experimental results demonstrate that the generated watermark images achieve a peak signal-to-noise ratio of up to 42dB, outperforming current mainstream baseline methods. Additionally, under various common image perturbation conditions, the model maintains stable watermark decoding accuracy, improving visual fidelity without sacrificing robustness. This effectively mitigates the "robustness-fidelity trade-off" problem and provides a feasible solution for copyright protection of generated content.?????关键词
人工智能生成内容，扩散模型，水印模型Key words
Artificial Intelligence Generated Content/ Diffusion Models/ Watermark Models引用本文复制引用
杜宝泽,李丽香.基于交错卷积的潜在空间水印模型[EB/OL].(2026-03-23)[2026-03-24].http://www.paper.edu.cn/releasepaper/content/202603-217.学科分类
计算技术、计算机技术
评论