Distributed-Backdoor

Requirements

Pytorch

Prepare the dataset

LOAN dataset:

• download the raw dataset into dir ./utils
• preprocess the dataset.

cd ./utils
bash process_loan_data.sh

Tiny-imagenet dataset:

• download the dataset into dir ./utils
• reformat the dataset.

cd ./utils
bash process_tiny_data.sh

Others:

MNIST and CIFAR10 will be automatically download.

Backdoor Attack

Modfiy the params.yaml file to conduct different experiment.

Attack A-M vs. A-S

Params:

• generation_epoch
  single-shot: [ different epochs for different attacker ]
  multi-shot: [ same epoch for attackers ]
• baseline
  single-shot: false; multi-shot: true
• eta
  single-shot: 0.1; multi-shot: 1
• poison_epochs
  single-shot: [ single epoch for each attacker ]
  multi-shot: [ multiple epochs for attackers ]

Distributed vs. Centralized

Params:

• adversary_list
  distributed attackers: [ multiple attackers ]
  centralized attacker: [ single attacker ]
• poison_epochs
  single-shot: X_poison_epochs
  multi-shot: 0_poison_epochs

Trigger Generation

Our method for trigger generation is incorporated in gen.py
Set is_generated: true to apply our scheme.

Reproduce experiments

We can use Visdom to monitor the training process.

python -m visdom.server -p 8098

Run experiments for different datasets:

python main.py --params utils/X.yaml

X = mnist_params, cifar_params,tiny_params or loan_params. Parameters can be changed in those yaml files to reproduce our experiments.