SecureFedYJ: a safe feature Gaussianization protocol for Federated Learning
Authors: Tanguy Marchand, Boris Muzellec, Constance Béguier, Jean Ogier du Terrail, Mathieu Andreux
NeurIPS 2022 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Quantitative experiments on real data demonstrate that, in addition to being secure, our approach reliably normalizes features across silos as well as if data were pooled, making it a viable approach for safe federated feature Gaussianization. |
| Researcher Affiliation | Industry | Tanguy Marchand Owkin Inc., New York, USA. EMAIL Boris Muzellec Owkin Inc., New York, USA. EMAIL Constance Beguier Jean Ogier du Terrail Owkin Inc., New York, USA. EMAIL Mathieu Andreux Owkin Inc., New York, USA. EMAIL |
| Pseudocode | Yes | Algorithm 1 EXPYJ ... Algorithm 2 EXPUPDATE ... Algorithm 3 SECUREFEDYJ |
| Open Source Code | No | The code of the experiments is not provided, but a detailed pseudo-code of the newly proposed algorithms are provided. |
| Open Datasets | Yes | All datasets used in this work are publicly available from the UCI machine learning repository [15]. ... [15] Dheeru Dua and Casey Graff. UCI machine learning repository, 2017. URL http://archive.ics.uci.edu/ml. |
| Dataset Splits | Yes | We evaluate each strategy using 5-fold cross-validation with 5 different seeds. |
| Hardware Specification | No | The experiment are not heavy and run easily on a personal computer, on a CPU |
| Software Dependencies | No | We implement SECUREFEDYJ in Python, using the MPy C library [42] based on Shamir Secret Sharing [44]. ... Standard implementations of the YJ transformation, in particular the scikit-learn implementation [36]... |
| Experiment Setup | Yes | We specified all the hyperparameters and the details of the numerical experiment in Appendix E. ... For the Cox PH model, we train it with L1 regularization (alpha = 0.5) using the CPHNN-L1 solver from the lifelines Python package. We select the learning rate among {1e-3, 1e-4, 1e-5} based on cross-validation on the training set. |