SSTQ: Privacy-Preserving Vector Quantization via Subsampled Stochastic TurboQuant
Original title:SSTQ:Privacy-Preserving Vector Quantization via Subsampled Stochastic TurboQuant
The paper introduces Subsampled Stochastic TurboQuant (SSTQ), a privacy-preserving vector quantization framework for distributed optimization. It combines overcomplete equal-norm tight frames, coordinate subsampling, and privacy-aware one-dimensional quantization. SSTQ provides a Flat Randomized Response variant and a Metric-Aware Laplace variant, with the latter targeting higher codebook bit widths. The authors derive optimal mean-squared-error scaling with ceil(log2 N)+b bits per client, where N=Theta(d), and propose a surrogate codebook objective that reduces codebook-dependent MSE scaling from O(4^b) to O(2^b). Experiments on federated CIFAR-10 and Fashion-MNIST report favorable utility and communication efficiency against established baselines.
Why it's worth reading
It is timely because it addresses the coupled constraints of local differential privacy, quantization error, and federated communication, with a claimed improvement in codebook-dependent error that may affect higher-bit deployments.