Sensitivity-Guided Contrastive Forgetting: Unified Label and Feature Unlearning in Vertical Federated Learning
Full conference-ready paper. See repository for code and results.
Please refer to the paper.md file for the full content. The complete paper with all sections (Abstract, Introduction, Related Work, Method, Experiments, Results, Discussion, Conclusion) plus appendices is available in this repository.
Quick Summary
Abstract
We propose UFUSC — the first framework to simultaneously perform label and feature unlearning in Vertical Federated Learning. Through contrastive forgetting loss, Lipschitz feature sensitivity minimization, and dual-variable certification, UFUSC achieves state-of-the-art forgetting (4.8-13.0% forget accuracy) while maintaining competitive utility across MNIST, Fashion-MNIST, and CIFAR-10.
Key Results
MNIST: Retain 85.45%, Forget 13.00%, MIA ASR 29.70%
Fashion-MNIST: Retain 70.18%, Forget 3.00%, MIA ASR 19.10%
CIFAR-10: Retain 50.96%, Forget 4.80%, MIA ASR 38.80%