Bo-Hsu Ke You-Zhe Xie Yu-Lun Liu Wei-Chen Chiu

Abstract

3D scene representation methods like Neural Radiance Fields (NeRF) and 3DGaussian Splatting (3DGS) have significantly advanced novel view synthesis. Asthese methods become prevalent, addressing their vulnerabilities becomescritical. We analyze 3DGS robustness against image-level poisoning attacks andpropose a novel density-guided poisoning method. Our method strategicallyinjects Gaussian points into low-density regions identified via Kernel DensityEstimation (KDE), embedding viewpoint-dependent illusory objects clearlyvisible from poisoned views while minimally affecting innocent views.Additionally, we introduce an adaptive noise strategy to disrupt multi-viewconsistency, further enhancing attack effectiveness. We propose a KDE-basedevaluation protocol to assess attack difficulty systematically, enablingobjective benchmarking for future research. Extensive experiments demonstrateour method's superior performance compared to state-of-the-art techniques.Project page: https://hentci.github.io/stealthattack/