Diffusion-Latent Invisible Watermarking for Proactive Deepfake Provenance Verification

Abstract

The rapid advancement of generative face manipulation techniques has exposed fundamental limitations of conventional passive deepfake detection systems, particularly under distribution shift, compression, and pose variation. In this paper, we propose diffusion-latent invisible watermarking as a proactive deepfake provenance verification mechanism, enabling reliable authentication of synthetic content through diffusion-latent provenance verification. Our approach embeds structured, imperceptible watermarks directly into the latent space of diffusion models, ensuring minimal visual distortion while preserving robust recoverability under post-processing operations. Unlike existing watermarking methods that are primarily evaluated on near-frontal facial imagery, our framework explicitly addresses non-frontal and side-face deepfake scenarios, where facial geometry and texture statistics differ substantially, and robustness is significantly challenged. We introduce a perceptual-robust latent reconstruction objective that enhances watermark stability under compression, geometric transformations, and learned distortions without sacrificing visual fidelity. To rigorously evaluate the proposed approach, we conduct comprehensive experiments on public datasets, featuring diverse facial identities, multiple state-of-the-art manipulation methods, and controlled pose variations. Results show that our method consistently surpasses prior diffusion-based watermarking baselines in both watermark detection and provenance verification reliability, especially under extreme pose variations and challenging post-processing conditions. These results highlight the effectiveness of proactive, diffusion-latent invisible watermarking as a practical and robust solution for trustworthy deployment of generative face technologies. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2027.