defense.ft_sam

class ft_sam[source]

Bases: defense

Enhancing Fine-Tuning Based Backdoor Defense with Sharpness-Aware Minimization

basic structure:

  1. config args, save_path, fix random seed

  2. load the backdoor attack data and backdoor test data

  3. load the backdoor model

  4. for each round sample a clean batch from given clean subset:

    1. do weight perturb to maximize L constrained by rho

    2. do outer minimization

  5. test the result and get ASR, ACC, RC

parser = argparse.ArgumentParser(description=sys.argv[0])
ft_sam.add_arguments(parser)
args = parser.parse_args()
ft_sam_method = ft_sam(args)
result = ft_sam_method.defense(args.result_file)

Note

@InProceedings{Zhu_2023_ICCV, author = {Zhu, Mingli and Wei, Shaokui and Shen, Li and Fan, Yanbo and Wu, Baoyuan}, title = {Enhancing Fine-Tuning Based Backdoor Defense with Sharpness-Aware Minimization}, booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)}, month = {October}, year = {2023}, pages = {4466-4477}}

Parameters:

  • args (baisc) – in the base class

  • rho (float) – Rho parameter for SAM.

  • adaptive (bool) – True if you want to use the Adaptive SAM.

  • label_smoothing (float) – Use 0.0 for no label smoothing.

  • rho_max (float) – max of Rho parameter for SAM.

  • rho_min (float) – min of Rho parameter for SAM.

  • alpha (float) – alpha parameter for gradient_decompose part of SAM.

  • checkpoint_path (str) – specify the checkpoint