Researchers introduce Privatar, a framework that offloads avatar reconstruction from headsets to untrusted local devices while safeguarding against data interception. The system leverages domain-specific knowledge of avatar reconstruction to achieve provably private offloading with minimal performance cost.

  • Horizontal Partitioning (HP) decomposes avatar reconstruction via BDCT in the frequency domain, keeping high-energy components on-device and offloading only low-energy subsets to reduce information leakage.
  • Distribution-Aware Minimal Perturbation (DAMP) minimizes noise based on each user's slowly changing expression distribution, reducing utility loss while maintaining formal local Differential Privacy guarantees.
  • On a Meta Quest Pro, Privatar supports 2.37x more concurrent users with only 6.5% higher reconstruction loss and 9% energy overhead compared to baselines.

Privatar provides both provable privacy guarantees and robustness against empirical and neural network-based attacks, offering a better throughput-loss Pareto frontier than quantization or sparsity methods.