Neural Refraction Fields for Image Verification

Image verification is an increasingly critical need as the capabilities for visual media tampering improve. We present an approach to image verification that embeds a private authenticity signature into an image based on physical refractive objects, which transform the image. To verify a protected image, we compare the image and a pixel-aligned reconstruction from the embedding to identify inconsistencies. Our approach is inspired by prior work, which physically places refractive objects in a scene before taking a photo. However, prior work is limited to simple refractions with analytical formulas and requires a slow per-scene neural radiance field optimization to get a reconstruction. Instead, our method trains a compact, scene-agnostic neural refraction field capable of modeling complex and more secure refractive geometries. Once trained, it enables instant, high-fidelity reconstruction for manipulation detection and localization.