attack.LIRA

class LIRA[source]

Bases: BadNet

LIRA: Learnable, Imperceptible and Robust Backdoor Attacks

basic structure:

  1. config args, save_path, fix random seed

  2. set the clean train data and clean test data

  3. set the device, model, generator, criterion, optimizer, training schedule.

  4. train the generator and classifier

  5. fix the generator, train the classifier

  6. save the attack result for defense

attack = LIRA()
attack.attack()

Note

@inproceedings{Doan2021lira, title = {LIRA: Learnable, Imperceptible and Robust Backdoor Attacks}, author = {Khoa D. Doan and Yingjie Lao and Weijie Zhao and Ping Li}, booktitle = {Proceedings of the IEEE International Conference on Computer Vision}, year = {2021}}

Parameters:

  • attack (string) – name of attack, use to match the transform and set the saving prefix of path.

  • attack_target (Int) – target class No. in all2one attack

  • attack_label_trans (str) – which type of label modification in backdoor attack

  • bd_yaml_path (string) – path for yaml file provide additional default attributes

  • random_crop (int) – random crop size

  • random_rotation (int) – random rotation degree

  • attack_model (string) – use which generator model

  • lr_atk (float) – learning rate for generator

  • eps (float) – epsilon for generated noise in train

  • test_eps (float) – epsilon for generated noise in test

  • alpha (float) – clean loss and poison loss ratio in train

  • test_alpha (float) – clean loss and poison loss ratio in test

  • finetune_lr (float) – learning rate for finetune

  • finetune_steplr_gamma (float) – gamma for finetune scheduler

  • finetune_steplr_milestones (list) – milestones for finetune scheduler

  • finetune_optimizer (string) – optimizer for finetune scheduler

  • both_train_epochs (int) – epochs for train both generator and classifier

  • train_epoch (int) – epochs for train classifier

  • verbose (bool) – verbosity of log

  • avoid_clsmodel_reinitialization (bool) – whether to avoid reinitialization of clsmodel during training

  • **kwargs (optional) – Additional attributes.