Poison-3DGS (review copy)

Anonymous data release accompanying a paper under double-blind review. For review purposes only. Please do not redistribute it or use it for any other purpose.

This review set contains 150 scenes (125 poisoned and 25 clean), built from 25 base scenes. The full benchmark of 1,353 scenes (1,253 poisoned and 100 clean) will be released after acceptance. The release contains poisoned training data and manipulated 3D Gaussian Splatting models.

Contents

benchmark/ holds, for each of the 25 base scenes, the clean scene and one poisoned variant of each of five attacks (StealthAttack, Poison-Splat, GaussTrap, 3D-GSW, GuardSplat).

benchmark/
├── clean/                                # 25 clean reference scenes
│   └── <dataset>/<scene>/
│       ├── images/                       # training images
│       ├── sparse/0/                     # COLMAP reconstruction
│       └── ply/
│           ├── iteration_7000/point_cloud.ply     # model after 7,000 training iterations
│           └── iteration_30000/point_cloud.ply    # final model
│
├── stealthattack/                        # illusory-object injection (25 scenes)
│   └── poisoned/<dataset>/<scene>/<injection>/    # single | single_mv | multi3
│       ├── images/                       # training images, including the trigger view(s)
│       ├── trigger_views/                # the trigger view(s) alone, same file names as in images/
│       ├── trigger.png / mask.png        # inserted object and its placement (single only)
│       ├── sparse/0/                     # COLMAP reconstruction; points3D.ply carries the injected points
│       └── s100_linear/                  # view noise sigma_0 = 100, linear decay
│           ├── config.yaml
│           └── ply/
│               ├── iteration_7000/point_cloud.ply
│               └── iteration_30000/point_cloud.ply
│
├── poisonsplat/                          # computation-cost attack (25 scenes)
│   └── poisoned/<dataset>/<scene>/eps16_100/      # eps = 16/255 on all views
│       ├── config.yaml
│       ├── images/                       # adversarially perturbed training images
│       ├── sparse/0/                     # COLMAP reconstruction of the clean scene
│       └── ply/
│           ├── iteration_7000/point_cloud.ply
│           └── iteration_30000/point_cloud.ply
│
├── gausstrap/                            # malicious content implanted by post-hoc fine-tuning (25 scenes)
│   └── poisoned/<dataset>/<scene>/full_d100_a5/
│       ├── config.yaml
│       └── ply/epoch_2500/point_cloud.ply
│
├── gsw/                                  # post-hoc watermark, 3D-GSW, 32-bit message (25 scenes)
│   └── poisoned/<dataset>/<scene>/gsw32_default/
│       ├── config.yaml
│       └── ply/epoch_8/point_cloud.ply
│
└── guardsplat/                           # post-hoc watermark, GuardSplat, 48-bit message (25 scenes)
    └── poisoned/<dataset>/<scene>/guard48_default/
        ├── config.yaml
        └── ply/iteration_30000/point_cloud.ply