Research map / Point clouds / Shape and surface completion
VVRec: Reconstruction Attacks on DL-based Volumetric Video Upstreaming via Latent Diffusion Model with Gamma Distribution
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes VVRec, a reconstruction attack that uses latent diffusion models with Gamma-distributed noise plus a point cloud refinement step to recover original point clouds from intermediate latents of DL-based volumetric video compressors.
Problem
Reconstruction attacks that recover original point clouds from intercepted intermediate representations of DL-based volumetric video upstreaming, where the attacker's training data distribution differs from the victim's.
Approach
Combines a shape latent diffusion model and a Gamma-noise latent point diffusion model (DDGM-style) with a VAE-style latent point generator and reconstruction decoder, then refines outputs via Shape-as-Points meshing and Poisson disk resampling to recover geometry and color.
Key results
- On MVUB-to-8iv2 at 0.4 bpp, VVRec reaches up to 64.70 dB p2plane-PSNR (M2) on Owlii-to-8iv2 for MSVoxelDNN-like settings, and averages about 18.91% M1 decline versus the Optimal decoder across seven victims.
- VVRec achieves up to 46.39% distortion reduction over baselines per the abstract, and up to 33.18% M1 improvement over the strongest baseline SP-GAN.
- Ablation: replacing Gamma with Gaussian noise lowers M2 by 6.61 dB; removing refinement lowers quality by 4.79%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use VVRec instead of SP-GAN | Reconstruction attack on PCGCv2 · MVUB attacker → 8iv2 victim, 0.4 bpp | p2p-PSNR (M1): +33.18% | Helped |
| Add SLDM and LPGDM (ablation) | VVRec-A · MVUB attacker → 8iv2 victim, PCGCv2 | p2plane-PSNR (M2) | Helped |
| Use Gamma instead of Gaussian noise in LPGDM (ablation) | VVRec-B · MVUB attacker → 8iv2 victim, PCGCv2 | p2plane-PSNR (M2): 6.61dB | Helped |
| Add the refinement module (ablation) | VVRec-C · MVUB attacker → 8iv2 victim, PCGCv2 | p2plane-PSNR (M2): 4.79% | Helped |
| Use VVRec instead of competing attack methods | Reconstruction attacks across datasets and victim models · Volumetric video datasets | p2p-PSNR (M1): up to 37.20% | Helped |
| Apply SI-Adv protection | VVRec reconstruction attack · Andrew@MVUB | p2plane-PSNR: less than 10% degradation | Hurt |
Surprising findings
- Existing protection SI-Adv, designed for attribute inference, only minimally reduces reconstruction accuracy (under 10% degradation).
- Attacker training on a subset-like dataset (Owlii, UVG-VPC) of 8iv2 yields better reconstruction than training on MVUB, since content similarity matters more than dataset size.
Limitations
- Diffusion sampling takes thousands of inference steps, making it slow for real-time attacks.
- Assumes the attacker has unlimited computation and no latency constraint.
- Attack may be countered by future defenses for DL-based streaming.
Experimental setup
Datasets: 8i Voxelized Full Bodies (8iv2) Owlii Dynamic Human Textured (Owlii) Microsoft Voxelized Upper Bodies (MVUB) UVG-VPC
Benchmarks: 8iv2 Owlii MVUB UVG-VPC
Metrics: p2p-PSNR p2plane-PSNR PCQM
Baselines: Supervised decoder SP-GAN VG-VAE PD-Flow Optimal (victim decoder)
Builds on: Denoising Diffusion Gamma Models PVCNN Shape-as-Points Latent diffusion models
Closest papers by meaning
- VRVVC: Variable-Rate NeRF-Based Volumetric Video Compression (AAAI 2025 · similarity 0.83)
- Understanding Reconstruction Attacks with the Neural Tangent Kernel and Dataset Distillation (ICLR 2024 · similarity 0.82)
- DRAG: Data Reconstruction Attack using Guided Diffusion (ICML 2025 · similarity 0.82)
- Blind Bitstream-corrupted Video Recovery via Metadata-guided Diffusion Model (CVPR 2025 · similarity 0.82)
- ConcreTizer: Model Inversion Attack via Occupancy Classification and Dispersion Control for 3D Point Cloud Restoration (ICLR 2025 · similarity 0.81)
- QUEEN: QUantized Efficient ENcoding of Dynamic Gaussians for Streaming Free-viewpoint Videos (NeurIPS 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.