Title: A Stage-by-Stage Analysis of Detectability in 3D Gaussian Splatting Poisoning

Abstract

While 3D Gaussian Splatting (3DGS) has quickly become a premier method for real-time novel view synthesis, it is increasingly susceptible to various poisoning attacks. These threats range from the injection of illusory objects and the amplification of computational costs to post-hoc model watermarking. Although the attack surface is expanding, current research predominantly emphasizes attack success rates, leaving defense mechanisms and detection strategies significantly underexplored.

A critical challenge—and opportunity—in detection stems from the multi-stage architecture of the 3DGS reconstruction pipeline, which generates heterogeneous intermediate representations. Forensic indicators of poisoning are inherently tied to specific pipeline stages; an intrusion at one point may only yield detectable signals at subsequent stages. This phenomenon necessitates a stage-wise approach to evaluating detectability, moving beyond traditional single-stage assessments.

To address this, we present Poison-3DGS, a benchmark designed for the stage-wise characterization of poisoning detection within 3DGS. This framework highlights stage-specific artifacts, such as multi-view imagery, geometric data, training dynamics, and Gaussian parameters, across a wide variety of scenes and attack vectors. Leveraging this benchmark, we perform a comprehensive analysis of detectability across the entire pipeline.

Our investigation yields several key findings: 1. Detectability fluctuates markedly across different stages, with no single stage consistently proving superior across all attack types. 2. Distinct attacks generate unique stage-specific forensic signals, meaning detection efficacy is heavily dependent on the timing of signal observation. 3. Signals emerging in later stages, particularly those related to training dynamics and Gaussian parameter statistics, offer robust cues that remain invisible during earlier phases.

Ultimately, this work establishes a principled benchmark and delivers the first systematic characterization of stage-dependent detectability in 3DGS, laying the groundwork for future advancements in building robust and reliable 3DGS systems.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC