Private Channel Perturbation for CNN Split Federated Learning

This repository contains the experimental scripts for a manuscript on client-private channel perturbation in CNN-based split federated learning. The method combines a private perturbation with a private bridge module so that the server-side model does not directly observe a client-aligned smashed representation.

The code was developed and run in Google Colab with the default Colab Python environment and a T4 GPU. The scripts are intentionally kept compact and correspond to the experiments reported in the paper.

Files

File	Purpose
sanity_shuffled_bridge.py	Main ResNet-18 split federated learning experiments across datasets.
sanity_shuffled_bridge_vgg11bn.py	Cross-architecture experiments with VGG11-BN.
sanity_shuffled_bridge_mobilenetv2.py	Cross-architecture experiments with MobileNetV2.
plot_comm_overhead_panel.py	Communication-overhead figure generation from finished JSON result files.

Experimental Setup

• Platform: Google Colab default environment
• GPU: NVIDIA T4
• Framework: PyTorch / torchvision
• Datasets: CIFAR-10, CIFAR-100, SVHN, and PathMNIST
• Backbones: ResNet-18, VGG11-BN, and MobileNetV2
• Split federated setting: 20 clients, 10 clients per round, Dirichlet non-IID partition with alpha = 0.3
• Main private mechanism: --defense shuffled_bridge --shuffled_bridge_cut A

The reconstruction attack results use two paper-aligned scenarios:

• Scenario A: private perturbation unknown to the attacker
• Scenario B: private perturbation known to the attacker

Baselines without private perturbation are reported as baseline_reconstruction.

Installation

In Colab, clone the repository and install the extra dependencies:

git clone https://github.com/EpiphanyMoment/split-sfl-channel-perturbation.git
cd split-sfl-channel-perturbation
pip install -r requirements.txt

Colab already provides PyTorch and torchvision. The requirements file also lists them for local reproduction.

Example Commands

Run the main ResNet-18 method on CIFAR-10:

python sanity_shuffled_bridge.py \
  --dataset cifar10 \
  --experiment_preset main20 \
  --defense shuffled_bridge \
  --shuffled_bridge_cut A \
  --stage both

Run baseline methods by changing --defense:

python sanity_shuffled_bridge.py --dataset cifar10 --experiment_preset main20 --defense none --stage both
python sanity_shuffled_bridge.py --dataset cifar10 --experiment_preset main20 --defense bottleneck --public_channels 24 --stage both
python sanity_shuffled_bridge.py --dataset cifar10 --experiment_preset main20 --defense pm --pm_group_size 4 --pm_init_prob 0.7 --stage both
python sanity_shuffled_bridge.py --dataset cifar10 --experiment_preset main20 --defense gaussian --noise_sigma 4 --stage both

Run the cross-architecture experiments:

python sanity_shuffled_bridge_vgg11bn.py --dataset cifar10 --defense shuffled_bridge --stage both
python sanity_shuffled_bridge_mobilenetv2.py --dataset cifar10 --defense shuffled_bridge --stage both

Generate the communication-overhead panel after the result JSON files are available:

python plot_comm_overhead_panel.py \
  --project_root . \
  --out_dir ./results_comm_overhead

Notes

• Dataset files, trained checkpoints, result JSON files, reconstruction images, and generated figures are not included in this repository.
• PathMNIST requires the medmnist package. For final PathMNIST runs, pass the same normalization statistics used in the paper experiments via --norm_json when available.
• LPIPS is optional at runtime. If the lpips package is unavailable, the scripts skip LPIPS and keep the other reconstruction metrics.

License

This project is released under the MIT License.