From Expansion to Retraction: Long-tailed Machine Unlearning via Boundary Manipulation

Min Chen, Weizhuo Gao, Chen Wang, Gaoyang Liu, Ahmed M. Abdelmoniem, Kai Peng

This is the official Pytorch implementation of our paper accept by KDD25.

Requirements

All codes are written by Python 3.7 with

• PyTorch = 1.13.1
• torchvision = 0.14.1
• numpy = 1.21.5

Preparing Datasets

Download the datasets CIFAR-10, CIFAR-100 to LTMU/data. The directory should look like

LTMU/data
├── CIFAR-100-python
└── CIFAR-10-batches-py

Training

When training the original model with varying imbalance ratios, one can adjust the imbalance_rate parameter. Specifically, it can be set to values within the range of [0.01, 0.02, 0.1].

for CIFAR-10-LT

python train.py --dataset cifar10 -a resnet18 --num_classes 10 --imbanlance_rate 0.01 --lr 0.01 --epochs 200 -b 64 --momentum 0.9 --weight_decay 5e-3

for CIFAR-100-LT

python train.py --dataset cifar100 -a resnet34 --num_classes 100 --imbanlance_rate 0.01 --lr 0.01 --epochs 200 -b 64 --momentum 0.9 --weight_decay 5e-3 

To perform retraining from scratch and obtain the retrain model of the specified class that you want to forget, you can add the parameter --forget_class forget_cls.

Unlearning

To execute our LTMU method, you can run the following code

python Unlearning.py --dataset cifar10 -a resnet18 --num_classes 10  --imbanlance_rate 0.01 --ori_path model_path --forget_class forget_cls --k number_of_k

To execute other baseline methods, you can run the corresponding code as follows. The method can be selected from the provided list which includes [BS, GA, RL, BT, FT, L1, GA_IA, GA_RFA, BS_IA, BS_RFA]

python Unlearning.py --dataset cifar10 -a resnet18 --num_classes 10  --imbanlance_rate 0.01 --ori_path model_path --forget_class forget_cls --unlearn_method method_name