padadmm

Code of the paper From Noisy Fixed-Point Iterations to Private ADMM for Centralized and Federated Learninghttps://arxiv.org/abs/2302.12559

Run main.py to redo the experiment of the figure 1 of the paper. Then run draw.py to generate the matplotlib figure.

The code is structured as follows:

• data.py generate the synthetic data
• conversion.py implements the conversion from eps delta differential privacy and Renyi DP
• dpadmm.py implements the DP-ADMM of the paper. Note that the algorithm is implement for the centralized, federated and decentralized version
• dpproxsgd.py implements the baseline with DP-Prox SGD
• gridsearch.py for tuning the parameters with the grid search
• lasso.py some utils function specific the Lasso objective

To cite the paper:

@article{DBLP:journals/corr/abs-2302-12559,
  author       = {Edwige Cyffers and
                  Aur{\'{e}}lien Bellet and
                  Debabrota Basu},
  title        = {From Noisy Fixed-Point Iterations to Private {ADMM} for Centralized
                  and Federated Learning},
  journal      = {CoRR},
  volume       = {abs/2302.12559},
  year         = {2023},
  url          = {https://doi.org/10.48550/arXiv.2302.12559},
  doi          = {10.48550/arXiv.2302.12559},
  eprinttype    = {arXiv},
  eprint       = {2302.12559},
  timestamp    = {Tue, 28 Feb 2023 14:02:05 +0100},
  biburl       = {https://dblp.org/rec/journals/corr/abs-2302-12559.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}