摘要：A fundamental challenge in data privacy is the privacy-utility trade-off: increasing privacy protection often reduces data utility. A key to mitigating this trade-off is to calculate the privacy budget (also known as privacy accounting) more accurately. In this talk, we introduce a general framework for tightly accounting for privacy budget in practice, leveraging statistical tools like hypothesis testing and mathematical tools from harmonic analysis and discrete math. We demonstrate its effectiveness in two applications: the 2020 US Census and fully decentralized federated learning algorithms.

For the 2020 US Census, tight privacy accounting is difficult because the budget depends on complicated discrete distributions. To address this, we propose a hybrid (numerical and analytical) privacy accountant that reduces the mean squared error of privatized contingency tables by 15.8% to 24.8% across geographical levels compared to the method used in current products.

In fully decentralized federated learning, privacy is amplified by data decentralization and random walks, which makes the budget depend on complicated mixture distributions that are hard to quantify. Using tools from hypothesis testing and Markov chain analysis, we developed a sharper privacy accountant that, in many cases, achieves significant improvements to the privacy budget (provides a >2x improvement).