ResSFL framework - improving SFL's resistance to model inversion attack

Official Repository for ResSFL (resistant split federated learning)

Requirement:

tensorboard>1.15

pytorch

torchvision

thop (pip install)

Code:

• MIA_torch.py: It implements the all utility functions of split learning,running MIA attacks and perform various defenses.
• main_MIA.py: Entry code to train a defensive model/vanilla model.
• main_test_MIA.py: Entry code to resume a trained model and perform MIA attack.

Proof Of Concepts:

Run scripts directly in ./reproduce/ to reproduce main results in the manuscript. You must have a GPU (CUDA-enabled) with >= 8 GB memory.

Recommend sequence to execute:

bash reproduce/ace_run_without_defense.sh
bash reproduce/ace_run_table6.sh

Comparison table:

bash reproduce/ace_run_table4.sh

Multi-client performance (with sampling, scale to 100 clients):

bash reproduce/ace_run_table2.sh
bash reproduce/ace_run_table2_sample.sh

List of additional experiments in supplementary material:

1. Performance against Optimization-based MI attack
2. Performance on clients with Non-iid data
3. Extra empirical evidence on successful resistance transfer

Cite the work:

@inproceedings{li2022ressfl,
  title={ResSFL: A Resistance Transfer Framework for Defending Model Inversion Attack in Split Federated Learning},
  author={Li, Jingtao and Rakin, Adnan Siraj and Chen, Xing and He, Zhezhi and Fan, Deliang and Chakrabarti, Chaitali},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={10194--10202},
  year={2022}
}