SSTQ: Privacy-Preserving Vector Quantization via Subsampled Stochastic TurboQuant
A novel approach for privacy-preserving distributed optimization, known as Subsampled Stochastic TurboQuant (SSTQ), has been introduced in a paper available on arXiv (2608.05127). This framework integrates overcomplete equal-norm tight frames, coordinate subsampling, and privacy-conscious one-dimensional quantization to ensure local differential privacy while keeping communication expenses low. It overcomes the drawbacks of current techniques like vqSGD, which are affected by dimension-dependent variance. SSTQ presents two versions: Flat Randomized Response and Metric-Aware Laplace, the latter being ideal for larger codebook bit-widths. The framework achieves optimal mean squared error scaling with only ceil(log2 N) + b bits per client, where N = Theta(d) denotes the frame size. A surrogate privacy-aware codebook objective enhances MSE scaling from O(4^b) to O(2^b). The authors of the paper are researchers, and it has been published on arXiv.
Key facts
- SSTQ is a new framework for privacy-preserving distributed optimization.
- It combines overcomplete equal-norm tight frames, coordinate subsampling, and privacy-aware one-dimensional quantization.
- SSTQ achieves local differential privacy while maintaining low communication cost.
- It includes two variants: Flat Randomized Response and Metric-Aware Laplace.
- SSTQ achieves optimal mean squared error scaling using only ceil(log2 N) + b bits per client.
- The frame size N = Theta(d).
- A surrogate privacy-aware codebook objective reduces MSE scaling from O(4^b) to O(2^b).
- The paper is available on arXiv with ID 2608.05127.
Entities
Institutions
- arXiv