Tracing the Use of Open-Source Training Datasets for Neural Radiance Field Models

Neural Radiance Fields (NeRF) have emerged as a pivotal method for 3D content representation. As NeRF technology advances, the ability to trace its training datasets becomes increasingly critical, highlighting the need for reliable dataset traceability mechanisms. While various approaches have been proposed to embed digital watermarks into NeRF models, these methods often necessitate modifications during the model’s training process to incorporate the watermark. Such alterations are impractical when the primary objective is to trace the training dataset itself, rather than the model. In this work, we introduce TraceNeRF, the first method designed to establish robust traceability of NeRF training datasets. This is achieved by embedding owner-specific binary messages into the datasets. In detail, we propose a novel hybrid frequency-spatial embedding framework, which combines a learnable spatially aware mask and discrete cosine transform. Furthermore, we incorporate a density-aware module together with a trainable band selector to enhance the robustness and persistence of watermark embedding. The experimental results demonstrate that our method outperforms other baseline methods in terms of availability, effectiveness, and robustness.