| Algorithm |
Type |
Debut Year |
Description |
| FedAvg (mcmahan2017communication, )
|
Classic |
2017 |
| the most typical FL algorithm. It adopts a process of distributing models, local training, uploading gradients, |
| and global aggregation. |
|
| FedProx (FedProx, )
|
Global Variable |
2020 |
| an improved version of FedAvg which enhances performance by adding a proximal term to the original loss |
| function and increasing the tolerance of the cloud server towards locally insufficient training gradients. |
|
| CluSamp (clusamp, )
|
Clustering |
2021 |
| an algorithm based on clustering, where clients are clustered and sampled based on sample size or model |
| similarity before training. |
|
| FedCross (fedcross, )
|
Multi-Model |
2024 |
| an algorithm specifically designed to address non-IID data in FL by cross-aggregating effectively and |
| distributing intermediate model based on a distance judgment and using the global model only for inference. |
|
|
Moon (moon, )
|
Contrastive
Learning-based
|
2021 |
an efficient FL algorithm based on contrastive learning. By introducing multiple representations to add a
model-contrastive loss during local training, Moon can control the training deviation of local models relative
to the global distribution.
|
| FedMut (fedmut, )
|
Mutation |
2024 |
| a method that mutates the global model based on gradient changes to generate multiple intermediate models |
| for the next round of training. Each intermediate model will be dispatched to the client for local training. |
| Finally, the global model converges to a flat region within the range of the mutation model. |
|
