class-wise

Worst-case forget set on class-wise unlearning

This is the official repository for Worst-case forget set on class-wise unlearning. The code structure of this project is adapted from the DP4TL codebase.

Requirements

You can install the necessary Python packages with:

pip install -r requirements.txt

We remark that to accelerate model training, this code repository is built based on FFCV and we refer its installation instructions to its official website. In this work, we build our argument system via fastargs, and we provide a revised version here. The installation of the latest fastargs is automatically handled by the command above.

Datasets

For ImageNet, we provide the preprocessed data (.beton) in this link. Please download the data and put them in the data folder. Replace the train_path and val_path of the dataset in the .json file within the configs with the path of the preprocessed data (.beton).

Scripts

In this section, we provide the instructions to reproduce the results in our paper.

1. Get the origin model. We first train the origin model (ResNet-18) on ImageNet using the following command:

   python src/experiment/imagenet_train_from_scratch.py --config-file configs/imagenet_train_from_scratch.json 

2. Replace the model_path of the blo in the .json file in the configs with the path of origin weight (.ckpt). Find the Worst-case forget set on class-wise unlearning. The selection weight will be saved at file/experiments/selection_worst_case.

   python src/experiment/class_wise_worst_case_mu.py --config-file configs/selection_worst_case.json --train.optimizer.lr 5e-5 --blo.w_lr 1e-4

3. Replace the training and testing of the indices in the .json file in the configs with the path of selection weight (.indices). Evaluate on the worst-case forget set.

   # Retrain
   python src/experiment/evaluation_retrain.py --config-file configs/evaluation_retrain.json --train.optimizer.lr {theta_lr}

   If you want to use approxiamte unlearning method(like FT, l1-sparse), you also should replace the model_path of the logging in the .json file in the configs with the path of origin weight (.ckpt).

   # l1-sparse
   python src/experiment/evaluation_ft.py --config-file configs/evaluation_retrain.json --train.optimizer.lr {theta_lr} --train.alpha {alpha}

   If indices are null, evaluation will be performed on the random forget set.