README.md

Dap-FL: Federated Learning flourishes by adaptive tuning and secure aggregation

Implemented based on Tensorflow

---

• The copyright of the code in this project belongs to the authors of the Dap-FL paper and can only be used for academic research. If you need to quote or reprint, please indicate the source and the original link.

• The final interpretation right of this copyright and disclaimer belongs to the authors.

---

Reference

• primal-dual ddpg from Accelerated Primal-Dual Policy Optimization for Safe Reinforcement Learning

Contents

dap_ddpg.py

define the basic ddpg model with an actor, a critic and a cost

ddpg_enhanced_second_train.py

execute FL training in dap

baseline_second_train.py

FL with fixed hyper-paraneters, i.e., FedAvg

mia.py

model inversation accacks

Experimental Settings

Machine Learning Tasks

• Logistic on MNIST

• CNN on MNIST

• ResNet on MNIST

• CNN on Fashion-MNIST

• CNN on FEMNIST

Adaptive FL Settings

• Dap-FL. (defined in ddpg_enhanced_second_train.py)

• Large. (defined in baseline_second_train.py)

• Small. (defined in baseline_second_train.py)

• DDPG-$\eta$. (defined in ddpg_enhanced_second_train.py by setting fixed training epoch)

• DDPG-$\alpha$. (defined in ddpg_enhanced_second_train.py by setting fixed learning rate)

• DDPG-client. (see https://ieeexplore.ieee.org/abstract/document/9372789/)

• DQN. (see https://ieeexplore.ieee.org/abstract/document/9244624/)