FedSkel

This repo is the implementation of FedSkel: Efficient Federated Learning on Heterogeneous Systems with Skeleton Gradients Update [CIKM'21].

Paper

FedSkel: Efficient Federated Learning on Heterogeneous Systems with Skeleton Gradients Update

Doi: 10.1145/3459637.3482107

Accuracy

We verified that FedSkel will not affect FL accuracy with PyTorch Framework. The core codes are in ./acc-pytorch. We adapted the FL framework of LG-FedAvg for FedSkel.

Speed Up

We measure the speedups of FedSkel on mobile devices with Caffe framework. The core codes are in ./speedup-caffe. We modified CONV layers of Caffe to enable gradients' pruning in the SetSkel and UpdateSkel processes during training.