FLAIR: Defense against Model Poisoning Attack in Federated Learning
Author: Atul Sharma
Requirement - PyTorch 1.8.1
The code when run downloads the dataset (MNIST and CIFAR-10) on its own, and it isnot available to setup the dataset separately.
Default run - "python main.py"
-- Runs FEDSGD aggregation with 10 benign clients on MNIST for 20 iterations
Example run - Running benign case followed by attack, followed by FLAIR's defense on MNIST
1) python main.py --gpu 0 --exp benign_fedsgd
2) python main.py --gpu 0 --byz_type full_trim --exp fulltrim_fedsgd
3) python main.py --gpu 0 --byz_type full_trim --aggregation flair --exp fulltrim_flair
The "main.py" file calls "aggregation.py" and "attack.py" to run the simulation of federated learning with the following command line arguments:
--dataset : mnist/ cifar10
--bias: controls the degree of non-iidness, 0.5 by default
--net : dnn for mnist, resnet18 for cifar10
--batch_size : 32 for mnist, 128 for cifar-10
--lr : learning rate - 0.01 for mnist, 0.1 for cifar10
--nworkers : 100 for mnist, 10 for cifar10
--nepochs : Number of training iterations, 500 for MNIST, 2000 for CIFAR-10
--gpu : 0 if CUDA is available, -1 otherwise
--nbyz : actual number of malicious workers (20% of nworkers)
--byz_type: benign/ full_trim/ full_krum/ adaptive_trim/ adaptive_krum/ shej_attack/ shej_agnostic
--aggregation : flair/ fedsgd/ krum/ trim (trimmed mean)/ median/ faba/ foolsgold/ fltrust
--cmax : workers the aggregator expects to be malicious, set to be equal to nbyz
--decay : decay parameter for FLAIR, default value = 0.99
--exp : name of the experiment given by the user
The code outputs a numpy vector - '<exp>_test_acc.npy' with the test accuracy values in every iteration, and '<exp>_model.pth' as the PyTorch model.
When the aggregation is FLAIR, two additional files are saved as outputs -
1)<exp>_FS.npy - Every row of this array represents one iteration, where every column is the flip-score of the respective client
Clients 0-19 are malicious and 20-99 are benign. Their behavior can be analyzed using this output file
2)<exp>_susp.npy - This contains the suspicion score of every client in a given iteration that was used to compute the reputation weights for aggregation in that iteration
For FoolsGold, FABA, and FLTrust, we output the corresponding client_list files that contains the reputation weight computed by those aggregation rules