CHESSFL: Clustering Hierarchical Embeddings for Semi-Supervised Federated Learning

Allen-Jasmin Farcas, Myungjin Lee, Ali Payani, Ramana Rao Kompella, Hugo Latapie, Radu Marculescu

Our proposed unified Semi-Supervised Hierarchical Federated Learning (SSHFL) evaluation framework. Edge servers (shown with orange border) have labeled data, while heterogeneous devices (shown with different colors, green, blue, red, purple) have their own local unlabeled data. Similar to classic HFL, we run steps 1 through 5 and evaluate the global model’s test accuracy after a given number of communication rounds.

Our proposed CHESSFL framework initiates training on the edge server with labeled data, utilizing our novel Sicilian Defense Loss. The feature embeddings derived from weakly-augmented labeled images are clustered, and the embedding cluster centroids are then transmitted to the client. Following the implementation of the proposed Pawn’s Progression Pseudo-Labeling the client starts training on unlabeled images using our novel Pawn Promotion Loss.

1. Create the virtual environment

conda create -n chessfl python==3.10
conda activate chessfl
pip3 install torch torchvision torchaudio paramiko scp tqdm pandas scipy scikit-learn

2. Prepare datasets

Run python generate_dataset.py to create the local datasets for all users.

3. Change paths

In utils/general_utils.py the function get_hw_info contains paths for the files. Change them accordingly.

You need to match the hw_type from get_hw_info(hw_type) with device_type from experiments/exp.bash.

4. Run experiments

Edit the experiment configuration experiments/exp.bash. Check the generate_configs.py for more details on the parameters used.

Run bash experiments/exp.bash to create the configuration.

Finally, run bash scripts/execute_exp.bash to run the experiment. You can edit this script to contain your desired experiments to run.

Citation

@inproceedings{farcas2024chessfl,
  title={CHESSFL: Clustering Hierarchical Embeddings for Semi-Supervised Federated Learning},
  author={Farcas, Allen-Jasmin and Lee, Myungjin and Payani, Ali and Kompella, Ramana Rao and Latapie, Hugo and Marculescu, Radu},
  booktitle={Proceedings of the 9th ACM/IEEE Conference on Internet of Things Design and Implementation},
  pages={122--133},
  year={2024}
}