On the Privacy-Robustness-Utility Trilemma in Distributed Learning

Authors: Youssef Allouah, Rachid Guerraoui, Nirupam Gupta, Rafael Pinot, John Stephan

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

Reproducibility Variable Result LLM Response
Research Type Experimental In Section E.1, we present our experimental setup. In Section E.2, we report our empirical results.
Researcher Affiliation Academia 1Ecole Polytechnique F ed erale de Lausanne (EPFL), Switzerland.
Pseudocode Yes Algorithm 1 SAFE-DSHB
Open Source Code No The code we use to launch the different experiments will be made available.
Open Datasets Yes We train a logistic regression model of d = 69 parameters on the academic Phishing5 dataset. (footnote 5: https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/)
Dataset Splits No The paper mentions using the Phishing dataset and shows 'Test accuracy' and 'Training Loss' but does not specify the dataset splits (e.g., train/validation/test percentages or counts) or the methodology used for splitting.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory, or cloud instance types) used for running its experiments.
Software Dependencies No We use Opacus (Yousefpour et al., 2021), a DP library for deep learning in Py Torch (Paszke et al., 2019).
Experiment Setup Yes We train the model using a fixed learning rate γ = 1 over a total of T = 400 learning steps. We set the clipping threshold C = 1 and the batch size b = 25. We run all algorithms, except DSGD, with momentum β = 0.99.