DWT-Augmented Adaptive Spread-Spectrum Watermarking in Diffusion Latents

Abstract

The proliferation of artificial intelligence-generated content (AIGC) has greatly boosted image synthesis efficiency, yet it also raises serious concerns over copyright infringement and content forgery. Existing watermarking methods generally show weak resistance to diffusion-based regeneration attacks, and struggle to balance embedding robustness with visual imperceptibility. To tackle these limitations, we present a blind watermarking framework that performs embedding operations directly within the latent representation space of a pre-trained diffusion model. Our approach adopts single-level Discrete Wavelet Transform (DWT) to decompose latent vectors into four separate subbands, and implants spread-spectrum watermark signals into high-frequency components to preserve overall visual quality. We further design a gradient-aware adaptive embedding mechanism that dynamically tunes watermark strength based on local feature gradients, which enhances robustness in texture-dense regions while suppressing visual artifacts in smooth image areas. Experiments on MS-COCO and DiffusionDB datasets verify that our method achieves favorable imperceptibility, maintains an ultra-low bit error rate (BER) under regeneration attacks, and outperforms state-of-the-art alternatives.