Hot PATE: Private Aggregation of Distributions for Diverse Task
Edith Cohen, Benjamin Cohen-Wang, Xin Lyu, Jelani Nelson, Tamas Sarlos, Uri Stemmer
TL;DR
Hot PATE extends Private Aggregation of Teacher Ensembles to diverse generative tasks by formalizing diversity-preserving ensemble samplers and introducing coordinated ensembles that transfer diversity without increasing privacy cost. The approach relies on coordinating teacher votes to produce histograms with high margins and bursty agreement, enabling both high utility (yield) and robust diversity transfer under threshold privacy and differential privacy. The authors provide formal definitions (tau-diversity, transfer, relevance) and prove that coordinations coupled with threshold or DP aggregators achieve strong privacy guarantees while preserving diversity. Empirically, Hot PATE significantly improves the privacy-utility trade-off in sequential text generation tasks, enabling privacy-preserving synthetic data generation and tunable-diversity outputs. The work demonstrates that diversity need not be sacrificed for privacy in open-ended generation and offers practical, plug-in mechanisms for private, diverse text generation.
Abstract
The Private Aggregation of Teacher Ensembles (PATE) framework enables privacy-preserving machine learning by aggregating responses from disjoint subsets of sensitive data. Adaptations of PATE to tasks with inherent output diversity such as text generation, where the desired output is a sample from a distribution, face a core tension: as diversity increases, samples from different teachers are less likely to agree, but lower agreement results in reduced utility for the same privacy requirements. Yet suppressing diversity to artificially increase agreement is undesirable, as it distorts the output of the underlying model, and thus reduces output quality. We propose Hot PATE, a variant of PATE designed for diverse generative settings. We formalize the notion of a diversity-preserving ensemble sampler and introduce an efficient sampler that provably transfers diversity without incurring additional privacy cost. Hot PATE requires only API access to proprietary models and can be used as a drop-in replacement for existing Cold PATE samplers. Our empirical evaluations corroborate and quantify the benefits, showing significant improvements in the privacy utility trade-off on evaluated in-context learning tasks, both in preserving diversity and in returning relevant responses.