GaussLock Uses Attribute-Space Traps to Defend 3D Gaussian Generative Models Against Unauthorized Fine-Tuning

The study proposes GaussLock, which jointly optimizes authorized distillation and attribute trap losses on position, scale, rotation, opacity, and color in parameter space to defend against unauthorized fine-tuning of 3D generative models. The traps compress the spatial distribution, distort geometric shapes, align rotation axes, and suppress primitive visibility, undermining the structural integrity of unauthorized reconstructions. Experiments on large-scale Gaussian models show that unauthorized reconstructions have higher LPIPS and lower PSNR, while authorized fine-tuning performance is largely maintained.