Awesome
Reprogrammable-FL: Improving Utility-Privacy Tradeoff in Federated Learning via Model Reprogramming
Reprogrammable - FL
The Structural Difference of Reprogrammable - FL with baselines
Environment
pip install opacus
conda install pytorch torchvision torchaudio pytorch-cuda=11.6 -c pytorch -c nvidia
pip install medmnist (optional for the medmnist dataset)
Training the centralized setting
python MR_Centralized.py ( For running the Reprogrammable Algorithm with DP SGD)
python BL_TS_Cent.py ( For running the Training from Scratch Algorithm with DP SGD)
python BL_FF_Cent.py ( For running the Fully Finetuning Algorithm with DP SGD)
python BL_PS_Cent.py ( For running the Partial Finetuning Algorithm with DP SGD)
Training the federated setting
python reprog_fl.py ( For running the Reprogrammable FL Algorithm with DP SGD)
python bl_scratch.py ( For running the Training from Scratch Algorithm with DP SGD)
python bl_ff.py ( For running the Fully Finetuning Algorithm with DP SGD)
python bl_pf.py ( For running the Partial Finetuning Algorithm with DP SGD)
References
Published in IEEE Conference on Security and Trustworthy Machine Learning 2023