RoPe-DW: Robust and Perceptual Dual Watermarking for Secure AI-Generated Multimedia Content

The proliferation of AI-generated multimedia content has posed significant challenges to multimedia retrieval systems, particularly in ensuring content authenticity and traceability. While latent diffusion models integrated with semantic watermarks (e.g., Gaussian Shading, Tree-Rings) offer promising solutions for content provenance, recent black-box semantic attacks have demonstrated that these watermarks can be effectively removed or forged using only watermarked images, without accessing model internals. This vulnerability undermines the reliability of watermark-based provenance analysis and limits the applicability of authenticity-aware multimedia retrieval systems. In this paper, we propose RoPe-DW, a novel latent diffusion framework with jointly optimized dual-watermark decoders that address this critical limitation. Our framework employs one-dimensional (1D) watermarks for provenance tracing and two-dimensional (2D) masks for tampering localization, enabling robust content identification even under adversarial manipulations. Extensive experiments demonstrate that RoPe-DW achieves significant improvements of 27% in robustness and 31.5% in perceptual attack resistance against black-box semantic attacks. This work integrates jointly optimized watermarking into latent diffusion models, enabling reliable authenticity and tamper detection, which provides a promising approach for enhancing the trustworthiness of AI-generated content in multimedia content analysis and retrieval applications.