Federated Label Unlearning

Official pytorch implementation of the paper: "Towards Privacy-Guaranteed Label Unlearning in Vertical Federated Learning: Few-Shot Forgetting Without Disclosure"

ICLR 2026 (OpenReview) | (ArXiv) |

(Released on February 08, 2026)

🚀 Introduction

We tackle label unlearning in Vertical Federated Learning (VFL), where labels are both necessary inputs and sensitive information. We introduce a representation-level manifold mixup to synthesize embeddings for unlearned and retained samples, providing stronger signals for efficient gradient-based forgetting and recovery. Our method removes label information while preserving utility via a lightweight recovery optimization, and scales across diverse datasets (e.g., MNIST, CIFAR-10/100, ModelNet, medical imaging, and Yahoo Answers).

Key Features:

• Efficiency: High-speed forgetting via manifold mixup.

• Utility Preservation: Lightweight recovery optimization to maintain model performance.

• Scalability: Validated on MNIST, CIFAR-10/100, ModelNet, Medical Imaging, and Yahoo Answers.

🛠️ Getting Start

1. Installation

Clone the repository and set up the Conda environment:

git clone https://github.com/bryanhx/Towards-Privacy-Guaranteed-Label-Unlearning-in-Vertical-Federated-Learning.git
cd Towards-Privacy-Guaranteed-Label-Unlearning-in-Vertical-Federated-Learning
conda env create -f environment.yml
conda activate unlearn

2. Dataset Preparation

Download and extract the datasets into a ./data directory at the project root:

Brain Tumor MRI

Yahoo Answer

ModelNet

🏃 Usage

Phase 1: Training the VFL Model

Train the baseline models or a retrained comparison model:

# Standard training
python main.py

Train a retrain model in 1 label unlearning scenario:

# Retraining
python main.py --mode=retrain

You may specify different data with --data=<data name>

For ModelNet, please use command below:

# Standard training
python main_modelnet.py --data=modelnet --num_classes=40

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

Phase 2: Unlearning

Note : Before running unlearning, update the torch.load() paths in the unlearning scripts (unlearn.py, unlearn_modelnet.py, etc.) to point to your saved model weights.

For all dataset except ModelNet, run following command below:

python unlearn.py

You may specify specific hyperparameters like data, model architecture, unlearn learning rate, unlearn epochs, and unlearn samples with --data, --model_type, --unlearn_lr. --unlearn_epochs and --unlearn_samples respectively.

For ModelNet:

python unlearn_modelnet.py --unlearn_method=LUV_modelnet --unlearn_lr=<unlearn learning rate> --unlearn_samples=<samples use for unlearning> --unlearn_epochs=<unlearning epochs>

Citation

@InProceedings{Hong_2026_ICLR,
    author    = {Gu, Hanlin and Tae, Hongxi and and Fan, Lixin and Chan, Chee Seng},
    title     = {Towards Privacy-Guaranteed Label Unlearning in Vertical Federated Learning: Few-Shot Forgetting Without Disclosure},
    booktitle = {The Fourteenth International Conference on Learning Representations (ICLR)},
    year      = {2026}
}

Feedback

Suggestions and opinions on this work (both positive and negative) are greatly welcomed. Please contact the author by sending an email to taehongxi55 at gmail.com or cs.chan at um.edu.my

License and Copyright

The project is open source under BSD-3 license (see the LICENSE file).

©2026 Universiti Malaya.