InvertedFontNet: Font Watermarking based on Perturbing Style Manifold

Nan Sun, Chenxin Zhao, Sijing Xie, Hefei Ling

Abstract: We introduce InvertedFontNet, a pioneering unsupervised font watermarking framework leveraging the font style manifold. Previous methodologies typically necessitate manual intervention or are confined by specialized font data and media. In contrast, our approach introduces an algorithm that exclusively relies on unlabeled font image data, enabling the embedding of extensive watermark information across diverse fonts. The algorithm strategically modifies font spatial structures by manipulating style manifolds, facilitating the embedding of watermarks via subtle glyph perturbations. Experimental results reveal the robustness of our algorithm against prevalent digital noise attacks, demonstrating superior detection accuracy compared to existing schemes.