Individual Privacy Accounting with Gaussian Differential Privacy

Authors: Antti Koskela, Marlon Tobaben, Antti Honkela

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

Reproducibility Variable Result LLM Response
Research Type Experimental We give experimental results that illustrate the benefits of replacing the RDP analysis with GDP accounting or with FFT based numerical accounting techniques.
Researcher Affiliation Collaboration Antti Koskela Nokia Bell Labs University of Helsinki antti.h.koskela@nokia-bell-labs.com Marlon Tobaben University of Helsinki marlon.tobaben@helsinki.fi Antti Honkela University of Helsinki antti.honkela@helsinki.fi
Pseudocode Yes Algorithm 1 Individual GDP Filter Algorithm
Open Source Code Yes Pythons codes needed for the experiments and plots is made publicly available (https://github.com/DPBayes/ individual-accounting-gdp).
Open Datasets Yes Our experiments use the MIMIC-III data set of pseudonymised health data by permission of the data providers... All the code related to MIMIC-III data set is publicly available (https://github.com/DPBayes/individual-accounting-gdp), as requested by Physionet (https://physionet.org/content/mimiciii/view-dua/1.4/).
Dataset Splits No While the paper mentions 'train set' and 'test set' and hyperparameter tuning, it does not explicitly specify the training, validation, and test dataset splits (e.g., percentages or sample counts) needed for reproduction.
Hardware Specification No The paper mentions general computing resources ('CSC IT Center for Science, Finland, and the Finnish Computing Competence Infrastructure (FCCI) for computational and data storage resources') but does not specify exact hardware details such as GPU/CPU models or memory used for experiments.
Software Dependencies No The paper mentions using 'Python' and 'opacus' for experiments, but does not provide specific version numbers for these or other relevant software dependencies, such as deep learning frameworks or libraries.
Experiment Setup Yes We train for 50 epochs with batch size 300, noise parameter σ = 2.0 and clipping constant C = 5.0. We train the model using DP-GD and opacus (Yousefpour et al.) with noise parameter σ 10.61 and determine the optimal clipping constant as C 0.79 in our training runs. We compute the budget B so that filtering starts after 50 epochs and set the maximum number of epochs to 100.