Robust 3D Watermarking for NeRF-Induced Modality Shifts

This study systematically addresses the issues of copyright infringement that have emerged with the introduction of neural radiance fields (NeRFs) by defining scenarios and applying and analyzing existing protective technologies tailored to each case. To formalize these threats, we introduce a threat model consisting of security requirements against modality shifts and an attacker model that classifies misuse based on the modalities held by the creator and attacker. This results in nine representative infringement cases that guide our experimental design. Our novel end-to-end watermarking approach is robust against modality changes caused by neural rendering. Independent decoders that vary according to the domain are proposed to extract messages from watermarked data effectively. In addition, a generalized neural rendering (GNR) module is utilized to optimize the various 3D models. These elements are integrated in an end-to-end manner, enhancing the interaction between different modalities and improving the overall performance. This study systematically examines the effectiveness of both the existing and proposed methods through experiments across different distribution modalities.