EncCluster

Scalable Functional Encryption in Federated Learning
through Weight Clustering and Probabilistic Filters

Vasileios Tsouvalas · Samaneh Mohammadi · Ali Balador · Tanir Ozcelebi · Francesco Flammini · Nirvana Meratnia

Abstract

EncCluster combines weight clustering, functional encryption and Binary Fuse filters for federated learning. Clients replace a full update with a compact centroid codebook, encrypt its values and encode where each centroid belongs. The server reconstructs the position mapping and recovers a sample-weighted aggregate through client-authorized functional keys. Local data stay on each client; clustering and filter precision control the accuracy–compression trade-off.

Setup

Create and activate the enccluster environment:

git clone https://github.com/Nousphera/EncCluster.git
cd EncCluster
python3.11 -m venv --prompt enccluster .venv
source .venv/bin/activate

From the activated enccluster environment, install the requirements:

python -m pip install -r requirements.txt

Data and Weights

Download, prepare and validate the datasets:

# One dataset
python scripts/prepare_data.py --dataset cifar10

# All paper datasets
python scripts/prepare_data.py

Datasets: cifar10 · cifar100.

Models: resnet20 · convnext · convmixer256_8 · mobilenet_v2. All train from scratch; convnext uses the paper's compact CIFAR architecture. No model download is needed.

Run Experiments

From the repository root, choose a method for the Flower simulation:

common=(--dataset cifar10 --model resnet20 --clients 30 --rounds 100
        --participation 1 --alpha 10 --local-epochs 1 --batch-size 64
        --lr 0.01 --kappa 128 --bpe 8 --precision 3 --device cuda)

# EncCluster (ours)
python scripts/run_experiment.py "${common[@]}" --method enccluster

# Weight clustering + Binary Fuse
python scripts/run_experiment.py "${common[@]}" --method bfuse

# Weight clustering
python scripts/run_experiment.py "${common[@]}" --method cluster

# FedAvg
python scripts/run_experiment.py "${common[@]}" --method fedavg

Note

--alpha 0.1 selects heterogeneous partitions. For 20% participation, use --participation 0.2 --rounds 300 (IID) or --rounds 500 (non-IID). --seed controls training; --partition-seed keeps client partitions fixed.

Results and final weights are saved under results/. --output chooses a new run name; --parallel-clients controls local concurrency. Use --device cpu or --device mps for local execution.

Aggregation

Each simulation round exchanges public keys, encrypted centroids, filters and mapping-bound authorization shares in bounded chunks. Floating model state, including BatchNorm statistics, stays encrypted; integer counters use the maximum. Results record ciphertext, mapping and authorization payload sizes separately.

Note

The protocol assumes authenticated peers and a semi-honest server. Mappings and released aggregates remain observable and can reveal centroid relationships. Authorization traffic is included explicitly; ciphertext compression alone does not establish the paper's total communication ratios.

Citation

@article{tsouvalas2025enccluster,
  title = {{EncCluster}: Scalable functional encryption in federated learning through weight clustering and probabilistic filters},
  author = {Tsouvalas, Vasileios and Mohammadi, Samaneh and Balador, Ali and Ozcelebi, Tanir and Flammini, Francesco and Meratnia, Nirvana},
  journal = {Pervasive and Mobile Computing},
  volume = {108},
  pages = {102021},
  year = {2025},
  publisher = {Elsevier},
  doi = {10.1016/j.pmcj.2025.102021},
  url = {https://doi.org/10.1016/j.pmcj.2025.102021}
}

License