Differentially Private Fair Learning

Authors: Matthew Jagielski, Michael Kearns, Jieming Mao, Alina Oprea, Aaron Roth, Saeed Sharifi -Malvajerdi, Jonathan Ullman

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental As a proof of concept, we empirically evaluate our two algorithms on a common fairness benchmark dataset: the Communities and Crime dataset from the UC Irvine Machine Learning Repository. ... Our main experimental goal is to obtain, for both algorithms, the Pareto frontier of error and fairness violation tradeoffs for different levels of differential privacy.
Researcher Affiliation Academia 1Northeastern University, Boston, MA, USA 2University of Pennsylvania, Philadelphia, PA, USA.
Pseudocode Yes Algorithm 1 ϵ-DP fair classification: DP-postprocessing; Subroutine 2 BESTϵ1 h; Algorithm 3 pϵ, δq-differentially private fair classification: DP-oracle-learner
Open Source Code No The paper does not provide any explicit statement about open-sourcing its code or include a link to a code repository.
Open Datasets Yes As a proof of concept, we empirically evaluate our two algorithms on a common fairness benchmark dataset: the Communities and Crime dataset from the UC Irvine Machine Learning Repository. We refer the reader to (Kearns et al., 2018a) for an outline of potential fairness concerns present in the dataset. We clean and preprocess the data identically to (Kearns et al., 2018a).
Dataset Splits No The paper mentions using a dataset of size m=2K but does not specify the training, validation, or test split percentages or counts used in the experiments.
Hardware Specification No The paper does not provide any specific hardware details such as GPU models, CPU types, or cloud computing resources used for the experiments.
Software Dependencies No The paper mentions using 'Logistic regression' and methods from other papers for implementation details, but does not specify software names with version numbers (e.g., Python, PyTorch, scikit-learn, etc.).
Experiment Setup Yes To elaborate, for a given setting of input parameters, we start with the target fairness violation bound γ 0 and then increase it over a rich pre-specified subset of r0, 1s while recording for each γ the error and the (realized) fairness violation of the classifier output by the algorithm. We take H to be the class of linear threshold functions, β 0.05, and δ 10 7.