Information Blackhole: Exploring Backdoor Mechanism in 3D Point Cloud Reconstruction
Researchers show point-cloud autoencoder backdoors leak source geometry and propose Gaussian constraints to control them.
The paper examines backdoor attacks on point-cloud autoencoders used for 3D reconstruction, arguing they differ from classifier backdoors because benign and poisoned samples share one latent distribution. Residual source geometry can leak into clean reconstructions and reduce attack controllability. The authors propose an Information Blackhole principle with Gaussian constraints and Adaptive Gaussian Matching to regularize poisoned latents. Experiments on ModelNet and ShapeNetPart show improved performance for several standard triggers.
- Point-cloud autoencoder backdoors differ from classifier backdoors because latents entangle data.
- Residual source geometry can leak into the clean reconstruction branch.
- Information Blackhole uses Gaussian constraints to block interfering source information.
- Adaptive Gaussian Matching regularizes poisoned latents to improve attack controllability.
- ModelNet and ShapeNetPart experiments improve several standard triggers.
- Products
- ModelNetShapeNetPart
From the original article
1 min read at arxiv.orgPoint cloud autoencoders are fundamental components for 3D world representation and support many safety-critical downstream applications. Existing studies have extensively investigated backdoor attacks on point cloud classification, whereas backdoor attacks against point cloud autoencoders remain largely unexplored.
By Zhifei Yang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.