Skip to content

Latest commit

 

History

History
21 lines (17 loc) · 781 Bytes

README.md

File metadata and controls

21 lines (17 loc) · 781 Bytes

MuGAN and zMuGAN

Machine unlearning techniques for zero-shot setups using GANs

The main script can be found in execute_unlearning_algorithms.py.

It takes as an input the following parameters:

  • model
  • weight_path
  • dataset
  • dataset_path is already downloaded
  • classes --> number of classes in this dataset
  • target_class --> target forget class id
  • gan_output --> use gan output vs original dataset
  • gan_dataset_size
  • learning_rate
  • lipschitz_std --> (only neeeded for JiT)
  • calc_ain ---> calculate Ain Score
  • method --> the intended unlearning technique:
    • the implemented options are: ('retrain','finetune', 'SCRUM' ,'UNSIR', 'negative_gradiant' , 'lipschitz', 'randomize_label', 'original', 'emmn', 'experimental_method')