Characterizing Detectability in 3DGS Poisoning: A Stage-wise Benchmark
Researchers have developed a new benchmark called Poison-3DGS to evaluate the detectability of poisoning attacks in 3D Gaussian Splatting (3DGS) systems. The benchmark analyzes how different stages of the 3DGS pipeline affect the visibility of these attacks. Their findings indicate that detection effectiveness varies significantly across stages, with later stages like training dynamics and Gaussian parameter statistics offering stronger forensic signals. AI
IMPACT Introduces a new benchmark for assessing the security of 3D Gaussian Splatting, potentially leading to more robust systems.