On Differentially Private Stochastic Convex Optimization with Heavy-tailed Data

Authors: Di Wang, Hanshen Xiao, Srinivas Devadas, Jinhui Xu

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments suggest that our algorithms can effectively deal with the challenges caused by data irregularity.Finally, we test our proposed aglorithms on both synthetic and real-world datasets. Experimental results are consistent with our theoretical claims and reveal the effectiveness of our algorithms in handling heavy-tailed datasets.
Researcher Affiliation Academia Di Wang * 1 2 Hanshen Xiao * 3 Srini Devadas 3 Jinhui Xu 1 1Department of Computer Science and Engineering, State University of New York at Buffalo, Buffalo, NY 2King Abdullah University of Science and Technology, Thuwal, Saudi Arabia 3CSAIL, MIT, Cambridge, MA.
Pseudocode Yes Algorithm 1 Sample-aggregate Framework (Nissim et al., 2007) Algorithm 2 Mechanism M in (Bun & Steinke, 2019) Algorithm 3 Heavy-tailed DP-SCO with known mean Algorithm 4 Heavy-tailed DP-SCO with known variance
Open Source Code Yes Due to the space limit, some definitions, all the proofs are relegated to the appendix in the Supplementary Material, which also includes the codes of experiments.
Open Datasets Yes For real-world data, we use the Adult dataset from the UCI Repository (Dua & Graff, 2017).
Dataset Splits No The paper only specifies a training and testing split ("28,000 amongst which are used as the training set and the rest are used for test") but does not explicitly mention a validation set or how it would be used for hyperparameter tuning or early stopping.
Hardware Specification No The paper does not provide specific hardware details (such as exact GPU/CPU models, processor types, or memory amounts) used for running its experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., library or solver names with version numbers like Python 3.8, CPLEX 12.4) needed to replicate the experiment.
Experiment Setup Yes For the privacy parameters, we will choose ϵ = {0.1, 0.5, 1} and δ = O( 1 /n). See Appendix for the selections of other parameters.