GaussTrap: Stealthy Backdoor Attacks on 3D Gaussian Splatting for Targeted Scene Misperception
GaussTrap is a novel poisoning attack targeting 3D Gaussian Splatting (3DGS) models to create scene-level confusion during inference. It injects malicious views at specific viewpoints to trigger harmful content while ensuring high-quality rendering remains intact in non-targeted views. This method addresses security flaws in 3DGS pipelines critical for safety applications.