English

Sample-Efficient Private Learning of Mixtures of Gaussians

Machine Learning 2024-11-05 v1 Data Structures and Algorithms Statistics Theory Machine Learning Statistics Theory

Abstract

We study the problem of learning mixtures of Gaussians with approximate differential privacy. We prove that roughly kd2+k1.5d1.75+k2dkd^2 + k^{1.5} d^{1.75} + k^2 d samples suffice to learn a mixture of kk arbitrary dd-dimensional Gaussians up to low total variation distance, with differential privacy. Our work improves over the previous best result [AAL24b] (which required roughly k2d4k^2 d^4 samples) and is provably optimal when dd is much larger than k2k^2. Moreover, we give the first optimal bound for privately learning mixtures of kk univariate (i.e., 11-dimensional) Gaussians. Importantly, we show that the sample complexity for privately learning mixtures of univariate Gaussians is linear in the number of components kk, whereas the previous best sample complexity [AAL21] was quadratic in kk. Our algorithms utilize various techniques, including the inverse sensitivity mechanism [AD20b, AD20a, HKMN23], sample compression for distributions [ABDH+20], and methods for bounding volumes of sumsets.

Keywords

Cite

@article{arxiv.2411.02298,
  title  = {Sample-Efficient Private Learning of Mixtures of Gaussians},
  author = {Hassan Ashtiani and Mahbod Majid and Shyam Narayanan},
  journal= {arXiv preprint arXiv:2411.02298},
  year   = {2024}
}

Comments

52 pages. To appear in Neural Information Processing Systems (NeurIPS), 2024