Saliency-Unlearning for Classification

This is the official repository for Saliency Unlearning for Clasification. The code structure of this project is adapted from the Sparse Unlearn codebase.

Requirements

pip install -r requirements.txt

Scripts

1. Get the origin model.

   python main_train.py --arch {model name} --dataset {dataset name} --epochs {epochs for training} --lr {learning rate for training} --save_dir {file to save the orgin model}
   

   A simple example for ResNet-18 on CIFAR-10.

   python main_train.py --arch resnet18 --dataset cifar10 --lr 0.1 --epochs 182
   

2. Generate Saliency Map

   python generate_mask.py --save_dir ${saliency_map_path} --model_path ${origin_model_path} --num_indexes_to_replace ${forgetting data amount} --unlearn_epochs 1
   

3. Unlearn

   • SalUn

   python main_random.py --unlearn RL --unlearn_epochs ${epochs for unlearning} --unlearn_lr ${learning rate for unlearning} --num_indexes_to_replace ${forgetting data amount} --model_path ${origin_model_path} --save_dir ${save_dir} --mask_path ${saliency_map_path}
   

   A simple example for ResNet-18 on CIFAR-10 to unlearn 10% data.

   python main_random.py --unlearn RL --unlearn_epochs 10 --unlearn_lr 0.013 --num_indexes_to_replace 4500 --model_path ${origin_model_path} --save_dir ${save_dir} --mask_path mask/with_0.5.pt
   

   To compute UA, we need to subtract the forget accuracy from 100 in the evaluation results. As for MIA, it corresponds to multiplying SVC_MIA_forget_efficacy['confidence'] by 100 in the evaluation results. For a detailed clarification on MIA, please refer to Appendix C.3 at the following link: https://arxiv.org/abs/2304.04934.

   • Retrain

   python main_forget.py --save_dir ${save_dir} --model_path ${origin_model_path} --unlearn retrain --num_indexes_to_replace ${forgetting data amount} --unlearn_epochs ${epochs for unlearning} --unlearn_lr ${learning rate for unlearning}
   

   • FT

   python main_forget.py --save_dir ${save_dir} --model_path ${origin_model_path} --unlearn FT --num_indexes_to_replace ${forgetting data amount} --unlearn_epochs ${epochs for unlearning} --unlearn_lr ${learning rate for unlearning}
   

   • GA

   python main_forget.py --save_dir ${save_dir} --model_path ${origin_model_path} --unlearn GA --num_indexes_to_replace 4500 --num_indexes_to_replace ${forgetting data amount} --unlearn_epochs ${epochs for unlearning} --unlearn_lr ${learning rate for unlearning}
   

   • IU

   python -u main_forget.py --save_dir ${save_dir} --model_path ${origin_model_path} --unlearn wfisher --num_indexes_to_replace ${forgetting data amount} --alpha ${alpha}
   

   • l1-sparse

   python -u main_forget.py --save_dir ${save_dir} --model_path ${origin_model_path} --unlearn FT_prune --num_indexes_to_replace ${forgetting data amount} --alpha ${alpha} --unlearn_epochs ${epochs for unlearning} --unlearn_lr ${learning rate for unlearning}