Practical Differentially Private Hyperparameter Tuning with Subsampling

Authors: Antti Koskela, Tejas D. Kulkarni

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We carry out experiments on several standard datasets, where we are able to improve upon the baseline tuning method by a clear margin.
Researcher Affiliation Industry Antti Koskela Nokia Bell Labs antti.h.koskela@nokia-bell-labs.com Tejas Kulkarni Nokia Bell Labs tejas.kulkarni@nokia-bell-labs.com
Pseudocode No The paper describes its method in numbered steps (e.g., 'Our method works as below:'), but does not provide a formally labeled 'Pseudocode' or 'Algorithm' block.
Open Source Code No The paper states 'For the implementation of DP-SGD and DP-Adam, we use the Opacus library (Yousefpour et al., 2021).' This refers to a third-party library used, not the open-source release of the specific methodology presented in this paper.
Open Datasets Yes We perform our evaluations on standard benchmark datasets for classification: CIFAR-10 (Krizhevsky and Hinton, 2009), MNIST (Le Cun et al., 1998), Fashion MNIST (Xiao et al., 2017) and IMDB (Maas et al., 2011).
Dataset Splits No In all of our experiments, we have a partitioning of the available data into train and test sets and we choose the best model based on the test accuracy. [...] the approach of making only two (train and test) partitions of available data instead of three (train, validation, and test) has been considered in many prior works.
Hardware Specification No The paper mentions 'a dedicated multi-GPU cluster' and 'two sets of gpus' but does not specify exact GPU models (e.g., NVIDIA A100) or other detailed hardware specifications like CPU models or memory.
Software Dependencies No The paper mentions using the 'Opacus library (Yousefpour et al., 2021)' and 'Ray Tune (Liaw et al., 2018)' but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes Full details of the experiments are given in Appendix A. [...] The learning rate grid size is either 9 or 10 and the grid values are specified in Table 2 (Appendix). We fix the subsampling ratio γ and the number of epochs to the values given in Table 1 (Appendix) for all models. [...] The number of trainable parameters and the hyperparameter grids are provided in Table 2 (Appendix B).