Differentially Private Statistical Inference through $\beta$-Divergence One Posterior Sampling

Authors: Jack E. Jewson, Sahra Ghalebikesabi, Chris C Holmes

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

Reproducibility Variable Result LLM Response
Research Type Experimental We provide extensive empirical evidence by reporting the performance of our model and four relevant baselines on ten different data sets, for two different tasks; additionally analysing their sensitivity in the number of samples and the available privacy budget.
Researcher Affiliation Academia Jack Jewson Department of Economics and Business Universitat Pompeu Fabra Barcelona, Spain jack.jewson@upf.edu Sahra Ghalebikesabi Department of Statistics University of Oxford Oxford, UK sahra.ghalebikesabi@univ.ox.ac.uk Chris Holmes The Alan Turing Institute Department of Statistics University of Oxford Oxford, UK chris.holmes@stats.ox.ac.uk
Pseudocode No The paper refers to algorithms from other works (e.g., 'Algorithm 1; [83]'), but does not include any pseudocode or algorithm blocks within its own content.
Open Source Code Yes Our code can be found at https://github.com/sghalebikesabi/beta-bayes-ops.
Open Datasets Yes The evaluations are conducted on simulated and UCI [23] data sets. For the latter, we have included the two classification data sets that were previously analysed in other applications of OPS (adult and abalone) [66, 81], in addition to other popular UCI data sets.
Dataset Splits No The paper states that 'validation splits' were used for hyperparameter tuning, but does not provide specific percentages or methodology for the training/validation split, only that test splits constitute 10% of the original data.
Hardware Specification Yes While the final experimental results can be run within approximately two hours on a single Intel(R) Xeon(R) Gold 5118 CPU @ 2.30GHz core, the complete compute needed for the final results, debugging runs, and sweeps amounts to around 11 days.
Software Dependencies No The paper mentions software like 'stan probabilistic programming language [16]' and 'sklearn', but does not provide specific version numbers for these or other key software components.
Experiment Setup Yes DPSGD is run for 14 + ϵ epochs, with clipping norm 1, batch size 100, and learning rate of 10 2.