Poison-3DGS: A Benchmark for 3D Gaussian Splatting Poisoning Detection

Poison-3DGS is a benchmark for evaluating detection methods against adversarial poisoning of 3D Gaussian Splatting assets. It was created by an anonymous user and covers 4 attack families across 37 scenes and 4 pipeline stages, as detailed in a NeurIPS 2026 paper.

Use Cases

• Benchmarking detection algorithms based on the described 4 attack families spanning integrity, availability, and steganography threats.
• Evaluating the robustness of 3D Gaussian Splatting pipelines based on attacks across 4 different pipeline stages.
• Researching adversarial machine learning in 3D vision based on the coverage of 37 different scenes.

Strengths

• Covers 4 distinct attack families, providing a multi-threat evaluation framework.
• Includes 37 different scenes, offering a variety of test cases.
• Evaluates attacks across 4 pipeline stages, allowing for stage-specific analysis.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment for large-scale training.

Provenance

Source

huggingface

Collection Method

Created as a benchmark for a NeurIPS 2026 paper.

Freshness

Last updated 2026-05-18 07:00:52; freshness should be verified.