MrNeRF (@janusch_patas)
2024-10-11 | โค๏ธ 24 | ๐ 2
Poison-splat: Computation Cost Attack on 3D Gaussian Splatting
TL;dr: Hack all the things.
Contributions: โข We reveal that the flexibility in model complexity of 3DGS can become a security backdoor, making it vulnerable to computation cost attack.
This vulnerability has been largely overlooked by the 3D vision and machine learning communities.
Our research shows that such attacks are feasible, potentially causing severe financial losses to 3D service providers.
โข We formulate the attack on 3D Gaussian Splatting as a data poisoning attack problem. To our best knowledge, there is no previous work investigating how to poison training data to increase the computation cost of machine learning systems.
โข We propose a novel attack algorithm, named Poison-splat, which significantly increases GPU memory consumption and slows down training procedure of 3DGS. We hope the community can recognize this vulnerability and develop more robust 3D Gaussian Splatting algorithms or defensive methods to mitigate such attacks.
๋ฏธ๋์ด
