A Primal-Dual Algorithm for Hybrid Federated Learning
Authors: Tom Overman, Garrett Blum, Diego Klabjan
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Furthermore, we provide experimental results that demonstrate the performance improvements of the algorithm over a commonly used method in federated learning, Fed Avg, and an existing hybrid FL algorithm, Hy FEM. |
| Researcher Affiliation | Academia | 1Department of Engineering Sciences and Applied Mathematics, Northwestern University 2Department of Mechanical Engineering, Northwestern University 3Department of Industrial Engineering and Management Sciences, Northwestern University tomoverman2025@u.northwestern.edu, garrettblum2024@u.northwestern.edu, d-klabjan@northwestern.edu |
| Pseudocode | Yes | Algorithm 1: Hy FDCA Initialize α0 = 0, w0 = 0, and ˆw0 = 0. Set ipk,i = 0 for every client k and i Ik. for t=1,2,...T do Given Kt, the subset of clients available in the given iteration Find Kt = {k : k Kt and k / Kt 1} Send enc(αt 0) to clients k Kt Primal Aggregation(Kt) Secure Inner Product(Kt) for all clients k Kt do αt k=Local Dual Method(αt 1 k , wt 1 k , x T i wt 1 0 ) Send all enc( γt |Bn| αt k) to server end for for n=1,2,...,N do enc(αt 0,n) = enc(αt 1 0,n ) + P b Bn enc( γt |Bn| αt b,n) end for Send enc(αt 0) to clients k Kt and clients decrypt Primal Aggregation(Kt) Secure Inner Product(Kt) end for |
| Open Source Code | No | The paper does not provide a direct link or explicit statement about open-sourcing the code for the described methodology. |
| Open Datasets | Yes | Three datasets are selected. MNIST is a database of handwritten digits (Deng 2012). News20 binary is a class-balanced two-class variant of the UCI 20 newsgroup dataset, a text classification dataset (Chang and Lin 2011). Finally, Covtype binary is a binarized dataset for predicting forest cover type from cartographic variables (Chang and Lin 2011). |
| Dataset Splits | No | The paper mentions "validation accuracy" and that "the value of λ that resulted in the highest validation accuracy is employed" for tuning, implying a validation set was used. However, it does not provide specific percentages, counts, or a detailed methodology for the dataset splits (training, validation, test). |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts, or detailed computer specifications) used for running its experiments. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers for its implementation (e.g., Python 3.8, PyTorch 1.9). |
| Experiment Setup | Yes | The regularization parameter, λ, is found by tuning via a centralized model where the value of λ that resulted in the highest validation accuracy is employed. The resulting choices of λ are λMNIST = 0.001, λNews20 = 1 10 5, and λcovtype = 5 10 5. ... For Hy FDCA, we only need to tune the number of inner iterations. |