Diffusion-Based Robust Watermarking with DPM-Solver++ Sampling for Facial Images Under Deepfake Manipulations

Abstract

The rapid progress of deepfake technology poses concerns over the authenticity of facial identity, the reliability of digital content, and social trust. Traditional watermarking techniques often are vulnerable to semantic changes due to face-swapping strategies that damage the durability of the watermark during identity-preserving changes. Specifically, the present paper brings forward a diffusion-based facial watermarking framework as WA-DPM Solver++, which adds saliency-guided adaptive embedding into the inverse diffusion process to alleviate this problem. Application of this technique involves global facial masking to preserve identity consistency and strong watermark insertion without an explicit spatial partitioning. The DPM-Solver++ was also used to speed up the sampling of diffusion and to further ensure embedding stability, leading to less computational overhead and faster convergence than with DDPM and DDIM. Experimental results on the CelebA-HQ dataset show that the proposed method presents significantly superior visual fidelity and robustness toward different image distortions and deepfake attacks while reaching a PSNR of 42.10 dB, SSIM at 0.9779, and BER uniformly low. Compared to DiffMark–DDIM and EditGuard, the proposed framework offers a better trade-off between efficiency, imperceptibility, and robustness. The findings suggest that combination of saliency-guided diffusion-based watermark embedding and DPM-Solver++ substantially boosts resistance against facial manipulation.