Megadose AI progress, ranked and analyzed.

SSTQ:Privacy-Preserving Vector Quantization via Subsampled Stochastic TurboQuant

· ArXiv · AI/CL/LG ·
SSTQ targets the privacy, variance, and bandwidth tradeoff in distributed vector quantization.

The paper proposes Subsampled Stochastic TurboQuant, combining tight frames, coordinate subsampling, and privacy-aware one-dimensional quantization. It describes two variants: Flat Randomized Response and Metric-Aware Laplace, with the latter aimed at higher bit-width codebooks. The authors claim optimal mean squared error scaling using `ceil(log2 N) + b` bits per client, with `N = Theta(d)`. They also report CIFAR-10 and Fashion-MNIST federated learning tests showing favorable utility and communication efficiency against baselines. ArXiv · AI/CL/LG's note

score 4

Categories: Research