DPFL-Robustnes

This repository provides the code for our ACM CCS 2023 paper titled Unraveling the Connections between Privacy and Certified Robustness in Federated Learning Against Poisoning Attacks.

Installation:

Our implementation was based on python=3.6.13, torch==1.10.1 and torchvision==0.11.2.

conda create -n dpflrobust python=3.6
conda activate dpflrobust
conda install pytorch==1.10.1 torchvision==0.11.2 cudatoolkit=11.3 -c pytorch
pip install -r requirements.txt

Datasets:

MNIST and CIFAR-10 will be download into the dir data

Run experiments:

1. DPFL training under adversarial users/instances

Train User-level DP clean models for certified prediction:

python main_userdp.py --config utils/mnist_params_ceracc.yaml  --fl_aggregation fedavg --dpfl max_model --n_runs 1

python main_userdp.py --config utils/cifar_params_ceracc.yaml  --fl_aggregation fedavg --dpfl max_model --n_runs 1

• --dpfl can be set as median_per_layer, max_per_layer, max_model,median_model for different clipping mechanisms.
• --n_runs is the number of runs for Monte Carlo Approximation

Train User-level DP models under $k$ adversarial users for certified attack inefficacy:

python main_userdp.py --config utils/mnist_params_cerattack.yaml --is_poison --fl_aggregation fedavg --dpfl max_model  --adv_method labelflip --scale_factor 100 --num_adv 10 --n_runs 1

python main_userdp.py --config utils/cifar_params_cerattack.yaml --is_poison --fl_aggregation fedavg --dpfl max_model  --adv_method backdoor --scale_factor 100 --num_adv 10 --n_runs 1

• --adv_method can be set as labelflip,backdoor
• --scale_factor is the factor for scaling the local update
• --num_adv is the number of adversarial users

Train Instance-level DP clean models for certified prediction:

python main_insdp.py --config utils/mnist_params_insdp_ceracc.yaml --n_runs 1 

python main_insdp.py --config utils/cifar_params_insdp_ceracc.yaml --n_runs 1

• --num_adv is the number of adversarial instances

Train Instance-level DP models under $k$ adversarial instances for certified attack inefficacy:

python main_insdp.py --config utils/mnist_params_insdp_cerattack.yaml --is_poison --adv_method labelflip --num_adv 5 --n_runs 1

python main_insdp.py --config utils/cifar_params_insdp_cerattack.yaml --is_poison --adv_method backdoor --num_adv 5 --n_runs 1

• --adv_method can be set as labelflip,backdoor

Hyperparameters can be changed according to our experiments setups (see appendix A.3 for paramerters details) in those yaml files to reproduce our experiments.