FedCVESA

This repository contains the slim code release for the FedCVESA v5 experiments used in the paper draft. The retained experiment setting is the 10-client Dirichlet non-IID split with alpha=0.5 on MNIST, Fashion-MNIST, and CIFAR-10.

What Is Included

• src/federated_main.py: main federated training and CVEA recovery entrypoint.
• src/models.py: only the v5 paper models:
  • CNNFashion_Enhanced for MNIST and Fashion-MNIST.
  • ResNet18Cifar for CIFAR-10.
• src/options.py: CLI options used by the v5 experiments.
• src/update.py, src/utils.py, src/sampling.py, src/attack_utils.py, src/plot.py: client training, partitioning, aggregation, attack, and plotting helpers.
• scripts_10clients/run_all_rank10_dirichlet05.sh: final v5 reproduction script.

Datasets, logs, .npy metrics, checkpoints, generated recovery images, paper figures, result summaries, and historical ablation scripts are intentionally not included.

Setup

Install dependencies from the repository root:

pip install -r requirments.txt

The dependency filename is kept as requirments.txt to match the original project. Datasets are downloaded by torchvision into data/ when an experiment first runs.

Quick Checks

Syntax check:

python -m compileall src

Small CPU sanity run:

python src/federated_main.py \
  --dataset=mnist \
  --model=cnn \
  --epochs=1 \
  --num_users=10 \
  --frac=1.0 \
  --iid=0 \
  --noniid_mode=dirichlet \
  --dirichlet_alpha=0.5 \
  --dirichlet_min_size=100 \
  --gama=0.5 \
  --num_steal=5 \
  --num_img_per_client=1 \
  --agg_mode=segmented \
  --attack_position_mode=spread

Reproduce v5

Preview the 57 jobs without running them:

DRY_RUN=1 bash scripts_10clients/run_all_rank10_dirichlet05.sh

Run all v5 jobs:

GPU_LIST="0 1 2 3 4 5 6 7" bash scripts_10clients/run_all_rank10_dirichlet05.sh

The script defaults to:

--num_users=10 --frac=1.0 --iid=0
--noniid_mode=dirichlet --dirichlet_alpha=0.5 --dirichlet_min_size=100
--gama_warmup_epochs=0 --lr_scheduler=cosine --min_lr=0.0001
MNIST/Fashion-MNIST: model=cnn, lr=0.01, local_ep=10, local_bs=16
CIFAR-10: model=resnet18_cifar, lr=0.03, local_ep=1, local_bs=64, cifar_crop_size=32, cifar_normalize=1

Outputs are generated under save/results, save/plots, save/objects, and scripts_10clients/logs_rank10_dirichlet05; these paths are ignored by git.