guangyaodou / conmu

Breaking the Trilemma of Privacy, Utility, Efficiency via Controllable Machine Unlearning

Controallble Machine Unlearning

Code instruction for Paper: Breaking the Trilemma of Privacy, Utility, Efficiency via Controllable Machine Unlearning (WWW 2024)

Introduction

Here is the Controllable Machine Unlearning (ConMU) repository. We conduct our experiments on CV benchmarks for both the random forgetting and class-wise forgetting.

Environment Download:

conda temp create -f conmu.yml

Random Forgetting

Here is the command to perform ConMU's random forgetting using ResNet-18 on CIFAR10 dataset.

python -u main.py --batch_size 128 --data CV_Data/cifar10 --dataset cifar10 --num_classes 10 --arch resnet18 --forget_percentage 0.2 
--save_dir ./dataset/resnet18 --epochs 100 --retrain_epoch 80 --lr 1e-3 --retrain_lr 1e-3 --prune_type rewind_lt --retain_filter_up 0.314 
--retain_filter_lower 0.314 --forget_filter_up 0.312 --forget_filter_lower 0.314 --num_noise 5 --further_train_lr 1e-2 --further_train_epoch 5 
--incompetent_epoch 3 --kl_weight 1.0

Parameters

• batch_size: batch size for the data
• data: location of the data corpus
• dataset: name of the dataset (cifar10, cifar100, svhn)
• num_classes: number of classes in the dataset
• arch: model architecture (Refer to cv_models/init.py).
• forget_percentage: percentage of data to forget
• save_dir: directory to save the model checkpoints
• epochs: number of epochs to train the original model
• retrain_epoch: number of epochs to train the gold model (retrained model using retained dataset)
• lr: learning rate for the original model
• retrain_lr: learning rate for the gold model
• num_noise: amount of noise (standard Gaussian) to add to the forgotten datasets
• further_train_lr: learning rate for the ConMU
• further_train_epoch: number of epochs to train the ConMU
• incompetent_epoch: number of epochs to train the unlearning proxy model

More details be be found at arg_parser.py.

Class-wise Forgetting

Here is the command to perform ConMU's class-wise forgetting using ResNet-18 on CIFAR10 dataset.

python -u main.py --batch_size 128 --data CV_Data/cifar10 --dataset cifar10 --num_classes 10 --arch resnet18 --forget_percentage 0.5 
--save_dir ./dataset/resnet18 --epochs 100 --retrain_epoch 80 --lr 1e-3 --retrain_lr 1e-3 --prune_type rewind_lt --retain_filter_up 0.314 
--retain_filter_lower 0.5 --forget_filter_up 0.5 --forget_filter_lower 0.5 --num_noise 5 --further_train_lr 1e-2 --further_train_epoch 5 
--incompetent_epoch 3 --kl_weight 0.5 --class_wise 5

Note that the the main differences between random forgetting and class-wise forgetting is that the --class_wise parameter is added to the command, which specifies the class number that we want to forget. Also, the default --forget_percentage is 0.5, which means that we forget 50% of the data from the specified class.