SplitFed: When Federated Learning Meets Split Learning
Authors: Chandra Thapa, Pathum Chamikara Mahawaga Arachchige, Seyit Camtepe, Lichao Sun8485-8493
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our analysis and empirical results demonstrate that (pure) SFL provides similar test accuracy and communication efficiency as SL while significantly decreasing its computation time per global epoch than in SL for multiple clients. |
| Researcher Affiliation | Collaboration | 1CSIRO Data61, Sydney, Australia 2Lehigh University, Bethlehem, Pennsylvania, USA {chandra.thapa, chamikara.arachchige, seyit.camtepe}@data61.csiro.au, lis221@lehigh.edu |
| Pseudocode | Yes | Algorithm 1: Splitfed Learning (SFL) and Algorithm 2: Client Update |
| Open Source Code | Yes | Some source codes are available at https://github.com/chandra2thapa/Split-Fed-When-Federated-Learning-Meets-Split-Learning. |
| Open Datasets | Yes | We use four public image datasets in our experiments, and these are summarized in Table 3. HAM10000 dataset is a medical dataset, i.e., the Human Against Machine with 10000 training images (Tschandl 2018). ... MNIST (Le Cun, Cortes, and Burges 2010), Fashion MNIST (Xiao, Rasul, and Vollgraf 2017), and CIFAR10 (Krizhevsky, Nair, and Hinton 2009) are standard datasets... |
| Dataset Splits | No | The paper lists 'Training samples' and 'Testing samples' in Table 3, but does not explicitly mention or detail any validation dataset splits or how validation was performed. |
| Hardware Specification | Yes | For quicker experiments and developments, we use the High-Performance Computing (HPC) platform that is built on Dell EMC s Power Edge platform with partner GPUs for computation and Infini Band networking. We run clients and servers on different computing nodes of the cluster provided by HPC. We request the following resources for one slurm job on HPC: 10GB of RAM, one GPU (Tesla P100-SXM2-16GB), one computing node with at most one task per node. The architecture of the nodes is x86 64. |
| Software Dependencies | Yes | All programs are written in python 3.7.2 using the Py Torch library (Py Torch 1.2.0). |
| Experiment Setup | Yes | The learning rate for Le Net is maintained at 0.004 and 0.0001 for the remainder of network architectures (Alex Net, Res Net, and VGG16). We consider the results under normal learning (centralized learning) as our benchmark. Table 5 summarizes our first result, where the observation window is 200 global epochs with one local epoch, batch size of 1024, and five clients for DCML. |