VVRec: Reconstruction Attacks on DL-based Volumetric Video Upstreaming via Latent Diffusion Model with Gamma Distribution

Rui Lu, Bihai Zhang, Dan Wang

Introduction

As immersive technologies like the Metaverse, Virtual Reality (VR), and autonomous driving transition into the mainstream, "volumetric video"—represented as massive collections of 3D points known as point clouds—is increasingly transmitted from users to cloud servers. This process, known as upstreaming, involves significant data traffic. To manage this, developers are shifting from traditional standards like MPEG toward Deep Learning (DL)-based compression techniques.

While DL-based compression offers high efficiency, it introduces a critical privacy vulnerability. The compressed "intermediate results" (latents) transmitted over the network often retain sufficient information to reconstruct the original sensitive input.

Overview of the VVRec attack scenario, showing how intercepted intermediate results can lead to the reconstruction of private user information.

The Challenge: Distribution Shift

A primary obstacle for attackers in this domain is "distribution shift." An attacker rarely possesses the exact dataset used to train a victim's specific encoder. VVRec (Volumetric Video Reconstruction) is designed to overcome this hurdle, demonstrating that even with different training data, high-fidelity reconstruction of a user's identity or environment is possible.

Methodology: The VVRec Pipeline

VVRec utilizes a multi-stage generative pipeline to "reverse" the compression process. The core innovation lies in its use of a diffusion model tailored for 3D data.

The Point Cloud Generation process within VVRec, highlighting the Latent Point Gamma Diffusion Model and the use of Gamma noise.

The pipeline consists of several key modules:

  1. Shape Latent Diffusion Model (SLDM): Processes intercepted "shape latents" through a diffusion process to denoise and stabilize the representation.
  2. Latent Point Generator (LPG): Converts abstract latent data into a structured point cloud format.
  3. Latent Point Gamma Diffusion Model (LPGDM): Unlike standard models that use Gaussian noise, VVRec employs a Gamma distribution. This choice better captures the sparse, non-uniform characteristics of high-dimensional point clouds, leading to higher capacity and faster convergence.
  4. Reconstruction Decoder (RD): A Variational Autoencoder (VAE)-based component that generates the initial 3D frame.
  5. Point Cloud Refinement (PCR): A post-processing step that converts raw points into a mesh and resamples them, significantly improving visual fidelity by smoothing surfaces and reducing noise.

Experimental Evaluation

The researchers evaluated VVRec using four industry-standard benchmarks: 8iv2, Owlii Dynamic Human, MVUB, and UVG-VPC. These datasets enabled "cross-dataset" attack testing, where the attack model is trained on one dataset (e.g., MVUB) to attack a victim using another (e.g., 8iv2).

Key Results

  • High Accuracy: VVRec achieved a reconstruction accuracy of 64.70dB (p2plane-PSNR), outperforming existing generative baselines like SP-GAN.
  • Distortion Reduction: The model demonstrated a 46.39% average reduction in distortion compared to state-of-the-art methods.
  • Resilience: The attack remained effective against common defenses like "SI-Adv" (adversarial noise injection), with less than a 10% drop in quality.
  • Visual Fidelity: Qualitative analysis showed that reconstructed figures were sharp enough to identify sensitive attributes such as facial features, gender, and race.

Discussion and Future Work

VVRec highlights that encryption alone (e.g., HTTPS) is insufficient if a malicious actor or cloud provider gains access to decrypted latents. The study's primary limitation is latency; the iterative nature of diffusion (1,000 steps) currently prevents real-time attacks. However, in many privacy-breach scenarios, time is not a limiting factor for the attacker.

Future research will likely focus on developing "privacy-preserving encoders" that can compress volumetric data without leaving structural signatures exploitable by generative models.

Summary updated

BibTeX
@inproceedings{lu2025vvrec,
  author = {Rui Lu and Bihai Zhang and Dan Wang},
  title = {{VVRec: Reconstruction Attacks on DL-based Volumetric Video Upstreaming via Latent Diffusion Model with Gamma Distribution}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i18.34106},
}
Figure