Sniper Backdoor

Code for the paper "Sniper Backdoor: Single Client Targeted Backdoor Attack in Federated Learning". SaTML'23.

Training the network and extracting the latent space:

IID

python main.py --lr 0.001 --dataname cifar100 --n_clients 10 --iid --dir <path> --n_epochs 23 --pretrained

Non-IID

python main.py --lr 0.001 --dataname cifar100 --n_clients 10 --iid --dir <path> --n_epochs 23 --pretrained

Generating synthetic data:

IID

pyhon synthetic_data.py --dataname cifar100 --dir <path> --n_clients 10 --iid

Non-IID

pyhon synthetic_data.py --dataname cifar100 --dir <path> --n_clients 10

Shadow training:

IID

python shadow_network.py --dataname cifar100 --n_epochs 23 --pretrained --dir <path> --fake_dir <results path> --iid

IID

python shadow_network.py --dataname cifar100 --n_epochs 23 --pretrained --dir <path> --fake_dir <results path>

Clients identification:

python client_identification.py --epochs 100 --n_clients 10 --dir <path> --dir_shadow <shadow network path> --dataname cifar100

Inject the Backdoor:

If we want to inject a dynamic backdoor we should use trojanzoo or other third party code. For static backdoors use:

python backdoor.py --dataname mnist --n_clients 5 --client_id <id> --source_label 1 --target_label 7