poisoning_enhancement

In this repository, code is for our paper [PoE: Poisoning Enhancement Through Label Smoothing in Federated Learning]

Installation

Install Python>=3.7.0 and Pytorch>=1.8.0

Usage

Prepare the dataset:

MNIST and CIFAR-10 dataset:

• MNIST and CIFAR will be automatically download

Code structures

• main.py, update,py, test.py, Fed_aggregation.py, enhancement.py: our FL framework
• main.py: entry point of the command line tool
• options.py: a parser of the FL configs, also assigns possible attackers given the configs
• poison_data.py: implement FL backdoor attack
• enhancement.py: code for label smoothing

Running Federated Learning tasks

Command line tools

python main.py --dataset mnist --model lenet5 --num_users 100 --rho 0.2 --epoch 150 --iid False --attack_start 4 --attack_methods CBA --attacker_list 2 5 7 9 --aggregation_methods EVE --detection_size 50 --gpu 0

Check out parser.py for the use of the arguments, most of them are self-explanatory.

• If you choose the backdoor method --attack_methods to be DBA, then you will also need to set the number of attackers in --attacker_list to 4 or a multiple of 4.
• If you choose the aggregation rule --aggregation_methods to be RLR, then you will also need to set the threshold in --robustLR_threshold.
• If --save_results is specified, the training results will be saved under ./results directory.