/thinking-two-moves-ahead

Primary LanguagePythonOtherNOASSERTION

Thinking Two Moves Ahead: Anticipating Other Users Improves Backdoor Attacks in Federated Learning

This code is the official implementation of Thinking Two Moves Ahead. Our implementation is based on Backdoor101.

If you have any questions, feel free to email Yuxin (ywen@umd.edu).

Dependencies

  • PyTorch => 1.11.*
  • torchvision >= 0.12.*
  • functorch >= 0.1.1

USAGE

non-IID CIFAR-10 baseline

python training.py --name baseline --fl-sample-dirichlet --fl-no-models 10 --fl-attack-freq 0 500 100 --loss-balance fixed --fixed-scales 1 --loss-tasks backdoor --update-method avg --fl-dp-clip 0.5 --fl-attacker-local-epochs 4 --fl-local-epochs 2 --lr 0.01 --params configs/cifar_fed.yaml --epochs 2000

non-IID CIFAR-10 anticipate

python training.py --name anticipate --anticipate --anticipate-steps 9 --fl-sample-dirichlet --fl-no-models 10 --fl-attack-freq 0 500 100 --loss-balance fixed --fixed-scales 1 --loss-tasks backdoor --update-method avg --fl-dp-clip 0.5 --fl-attacker-local-epochs 4 --fl-local-epochs 2 --lr 0.01 --params configs/cifar_fed.yaml --epochs 2000

Results

You can view the results through running tensorboard --logdir runs.