This is an official implementation of the paper LUV:A Few-shot Label Unlearning in Vertical Federated Learning

Getting Start

Environment

Create a conda environment with the following LUV_requirement.txt file

conda create -n <environment-name> --file LUV_requirement.txt

Datasets

Download the following datasets from the link provided. Place the datasets in the .\data directory.

Brain Tumor MRI: https://drive.google.com/drive/folders/1gFVOAGlUh-sCl-wbDzzrM9G_2UwtMCHB?usp=sharing

Yahoo Answer : https://drive.google.com/drive/folders/1Frwb-ozdsDCSwUbGKuXsj5bCbd3hIp8K?usp=sharing

Commands to train VFL model:

CIFAR10 Resnet18

Train Full Model:

python main.py

Train a retrain model in 1 label unlearning scenario:

python main.py --mode=retrain

Train a retrain model in 2 labels unlearning scenario:

python main.py --mode=retrain --unlearn_class_num=2

Train a retrain model in 4 labels unlearning scenario:

python main.py --mode=retrain --unlearn_class_num=4

MNIST Resnet18:

Train Full Model:

python main.py --data=mnist

Train a retrain model in 1 label unlearning scenario:

python main.py --data=mnist --mode=retrain

Train a retrain model in 2 labels unlearning scenario:

python main.py --data=mnist --mode=retrain --unlearn_class_num=2

Train a retrain model in 4 labels unlearning scenario:

python main.py --data=mnist --mode=retrain --unlearn_class_num=4

CIFAR100 Resnet18:

Train Full Model:

python main.py --data=cifar100 --num_classes=100

Train a retrain model in 1 label unlearning scenario:

python main.py --data=cifar100 --num_classes=100 --mode=retrain

Train a retrain model in 2 labels unlearning scenario:

python main.py --data=cifar100 --num_classes=100 --mode=retrain --unlearn_class_num=2

Train a retrain model in 4 labels unlearning scenario:

python main.py --data=cifar100 --num_classes=100 --mode=retrain --unlearn_class_num=4

Yahoo Answer MixText

Train Full Model:

python main.py --data=yahoo --model_type=mixtext --epochs=30

Train a retrain model:

python main.py --data=yahoo --model_type=mixtext --epochs=20 --mode=retrain --unlearn_class=6

ModelNet Resnet18:

Train Full Model:

python main_modelnet.py --data=modelnet --num_classes=40

Train a retrain model in 1 label unlearning scenario:

python main_modelnet.py --data=modelnet --num_classes=40 --mode=retrain

Brain MRI Resnet18:

Train Full Model:

python main.py --data=mri --num_classes=4

Train a retrain model in 1 label unlearning scenario:

python main.py --data=mri --num_classes=4 --mode=retrain --unlearn_class=2

CIFAR10 VGG16

Train Full Model:

python main.py --model_type=vgg16

Train a retrain model in 1 label unlearning scenario:

python main.py --mode=retrain --model_type=vgg16

Train a retrain model in 2 labels unlearning scenario:

python main.py --mode=retrain --unlearn_class_num=2 --model_type=vgg16

Train a retrain model in 4 labels unlearning scenario:

python main.py --mode=retrain --unlearn_class_num=4 --model_type=vgg16

CIFAR100 VGG16:

Train Full Model:

python main.py --data=cifar100 --num_classes=100 --model_type=vgg16

Train a retrain model in 1 label unlearning scenario:

python main.py --data=cifar100 --num_classes=100 --mode=retrain --model_type=vgg16

Train a retrain model in 2 labels unlearning scenario:

python main.py --data=cifar100 --num_classes=100 --mode=retrain --unlearn_class_num=2 --model_type=vgg16

Train a retrain model in 4 labels unlearning scenario:

python main.py --data=cifar100 --num_classes=100 --mode=retrain --unlearn_class_num=4 --model_type=vgg16

Before running the command for unlearning, change the saved model path directory in the torch.load() code from the unlearn python file.

Command for unlearning

Before running the unlearning Python files, ensure you update the model path in the torch.load() code to point to your saved directory in the following files: unlearn.py, unlearn_modelnet.py, unlearn_2labels.py, and unlearn_4labels.py.

1 label unlearning

MNIST Resnet18:

python unlearn.py --data=mnist --model_type=resnet18 --unlearn_lr=2e-7 --unlearn_epochs=10 --unlearn_samples=40

CIFAR10 Resnet18:

python unlearn.py

CIFAR100 Resnet18:

python unlearn.py --data=cifar100 --model_type=resnet18 --unlearn_lr=5e-7 --unlearn_epochs=7 --unlearn_samples=30

ModelNet Resnet18:

python unlearn_modelnet.py --unlearn_method=LUV_modelnet --unlearn_lr=5e-07 --unlearn_samples=30 --unlearn_epochs=4

Brain Tumor MRI Resnet18:

python unlearn.py --data=mri --model_type=resnet18 --unlearn_method=LUV --unlearn_epochs=4 --unlearn_samples=15 --unlearn_lr=6e-6 --unlearn_class=2

Yahoo Answer MixText:

python unlearn.py --data=yahoo --model_type=mixtext --unlearn_method=LUV --unlearn_class=6 --unlearn_lr=7e-07 --unlearn_samples=28 --unlearn_epochs=4

2 labels unlearning

CIFAR10 Resnet18:

python unlearn_2labels.py --data=cifar10 --model_type=resnet18 --unlearn_lr=1e-6 --unlearn_epochs=15 --unlearn_samples=40

CIFAR100 Resnet18:

python unlearn_2labels.py --data=cifar100 --model_type=resnet18 --unlearn_lr=9e-7 --unlearn_epochs=10 --unlearn_samples=20

CIFAR10 VGG16:

python unlearn_2labels.py --data=cifar10 --model_type=vgg16 --unlearn_lr=1e-6 --unlearn_epochs=15 --unlearn_samples=40

CIFAR100 VGG16:

python unlearn_2labels.py --data=cifar100 --model_type=vgg16 --unlearn_lr=9e-7 --unlearn_epochs=5 --unlearn_samples=20

4 labels unlearning

CIFAR100 Resnet18:

python unlearn_4labels.py --data=cifar100 --model_type=resnet18 --unlearn_lr=3e-6 --unlearn_epochs=20 --unlearn_samples=15

CIFAR100 VGG16:

python unlearn_4labels.py --data=cifar100 --model_type=vgg16 --unlearn_lr=1e-6 --unlearn_epochs=15 --unlearn_samples=15