Comparison Of Two Mean Vectors Under Differential Privacy For High-Dimensional Data

Caizhu Huang, Di Wang, Yan Hu, Nicola SartoriProceedings of the International Conference on Statistics, Theory and Applications (ICSTA ...) 2022 · DOI

Abstract

The multivariate hypothesis testing problem is a more interesting task of the statistical inference for high-dimensional data nowadays, in which the dimension of the observation vectors is diverging and could even be larger than the sample size. However, in many applications of multivariate hypotheses problems, the data are highly sensitive and require privacy protection. Here we consider a private non-parametric projection test for the comparison of the high-dimensional multivariate mean vectors that guarantees strong differential privacy. The empirical evidence shows that the non-parametric projection test under differential privacy gives accurate inference under the null hypothesis and a higher power under the local alternative hypothesis.

Introduction

In modern data analysis, particularly within genomics, finance, and social sciences, researchers frequently encounter the "Large $p$, Small $n$" problem. This occurs when the number of features ($p$) significantly exceeds the number of observations ($n$). Traditional statistical tools, such as Hotelling’s $T^2$ test, fail in these high-dimensional settings because the required covariance matrices become singular and mathematically intractable.

Furthermore, the sensitivity of such data—often containing personal health or financial records—demands robust privacy protections. Differential Privacy (DP) has emerged as the gold standard for such protections, but applying it to high-dimensional hypothesis testing without destroying the statistical utility of the data remains a significant challenge. This paper presents a non-parametric projection test designed to maintain high statistical power while ensuring strong privacy guarantees.

Methodology

The proposed approach integrates three core components: a projection-based statistic, the Laplace mechanism for Differential Privacy, and non-parametric permutation methods.

The Projection Statistic ($T_{PT}$)

Rather than analyzing all dimensions simultaneously, which introduces excessive noise in high-dimensional space, the test utilizes a projection term. By projecting the data onto a specific linear coefficient vector ($a$), the method highlights the separation between group means. The effectiveness of the test is highly dependent on the alignment of this vector with the actual difference between groups.

Differential Privacy via Laplace Mechanism

To protect individual privacy, the authors apply the Laplace mechanism. Noise is carefully added to the sample means and the trace of the covariance matrix. This ensures that the inclusion or exclusion of any single individual's data point does not significantly alter the test outcome, satisfying the requirements of $\epsilon$-differential privacy.

Non-Parametric Permutation Testing

To avoid making restrictive assumptions about the data's distribution (such as normality), the method employs permutation testing. By repeatedly shuffling group labels to generate a "null distribution," the researchers can calculate accurate p-values even for non-normal or unknown data distributions.

Experimental Evaluation

The method was evaluated using synthetic datasets generated from $p$-variate Uniform distributions ($U(-\sqrt{3}, \sqrt{3})$). Simulations covered dimensions ($p$) from 20 to 200 and sample sizes ($n$) as small as 5 per group. Two privacy budgets were tested: $\epsilon = 1/20$ (strong privacy) and $\epsilon = 1/4$ (moderate privacy).

Type I Error Stability

A critical requirement for any statistical test is maintaining a stable Type I error rate—the probability of incorrectly rejecting a true null hypothesis. As shown in Figure 1, the proposed private tests (npPT1, npPT2, npPT3) and the baseline (npBS) consistently maintained empirical Type I error rates near the nominal 0.05 level across all tested dimensions.

Empirical Type I error rates across varying dimensions ($p$) and privacy budgets ($\epsilon$).

Statistical Power

The primary advantage of the projection-based approach is its superior power under local alternative hypotheses. When a real difference exists between group means (quantified by $\delta$), the projection test $npPT3$ significantly outperforms the standard non-parametric baseline ($npBS$).

Statistical power comparison under local alternative hypotheses for varying difference levels ($\delta$).

As illustrated in Figure 2, at a difference level of $\delta = 9$, the $npPT3$ method achieved a power of approximately 0.325, while the baseline $npBS$ lagged at 0.095. This demonstrates that the projection method can detect subtle differences that standard methods miss, even after the addition of privacy-preserving noise.

Discussion and Conclusion

The results indicate that privacy does not have to come at the cost of statistical accuracy. The non-parametric projection test remains robust against high dimensionality and provides a reliable tool for sensitive data analysis.

However, some challenges remain. The permutation approach is computationally intensive, which may limit its application for extremely large sample sizes. Additionally, the test's power is sensitive to the choice of the projection vector $a$. Future work will focus on optimizing this selection and developing formal asymptotic theories to describe the test's behavior as $n$ and $p$ approach infinity.

Summary updated

Abstract

The multivariate hypothesis testing problem is a more interesting task of the statistical inference for high-dimensional data nowadays, in which the dimension of the observation vectors is diverging and could even be larger than the sample size. However, in many applications of multivariate hypotheses problems, the data are highly sensitive and require privacy protection. Here we consider a private non-parametric projection test for the comparison of the high-dimensional multivariate mean vectors that guarantees strong differential privacy. The empirical evidence shows that the non-parametric projection test under differential privacy gives accurate inference under the null hypothesis and a higher power under the local alternative hypothesis.
BibTeX
@inproceedings{huang2022comparison,
  author = {Caizhu Huang and Di Wang and Yan Hu and Nicola Sartori},
  title = {{Comparison Of Two Mean Vectors Under Differential Privacy For High-Dimensional Data}},
  booktitle = {Proceedings of the International Conference on Statistics, Theory and Applications (ICSTA ...)},
  year = {2022},
  doi = {10.11159/icsta22.167},
}
Figure