Answers are grounded in this paper and cite their sources.
With the rapid advancement of AI-generated content (AIGC) technologies, novel attack paradigms such as image generative tampering and resampling have posed severe challenges to existing robust watermarking algorithms. Current deep learning-based watermarking schemes generally suffer from the lack of reversibility and weak resistance against generative attacks, while traditional moment-based watermarking approaches fail to exploit the global semantic features of images, making them vulnerable to semantic-level tampering attacks. To address these issues, this paper proposes a synergistic robust reversible image watermarking scheme that fuses transformer semantic enhancement with invariant moment fusion, filling the critical research gap where no existing robust reversible watermarking method can effectively resist AIGC generative attacks. The scheme adopts a dual-branch feature fusion architecture combined with a two-stage reversible embedding mechanism. First, a vision transformer (ViT) extracts global semantic features, employing multi-head attention to precisely localize semantic core regions and generate robust representations. Second, the global semantic features are multi-scale deeply fused with geometric invariant features from Polar Harmonic Fourier Moments (PHFMs), constructing a fused feature space with both empirically observed semantic stability and mathematically guaranteed geometric invariance. The robust watermark is embedded into this fused feature space, significantly enhancing the watermark’s resistance against generative attacks and semantic tampering attacks. Third, an adaptive quantization index modulation algorithm ensures embedding imperceptibility and extraction stability. Finally, the integer wavelet transform losslessly embeds auxiliary recovery information, enabling complete original image restoration under attack-free scenarios. The experimental results demonstrate lossless image recovery and 100% accurate watermark extraction under no-attack conditions, maintaining a peak signal-to-noise ratio (PSNR) above 44 dB. Furthermore, the scheme demonstrates notable resilience against generative attacks—including AIGC regeneration and style transfer—as well as traditional geometric and signal processing attacks. Comparative evaluations show that the extraction accuracy yields substantial improvements over several existing state-of-the-art algorithms, providing a viable solution for image copyright protection in the AIGC era.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: