File size: 2,640 Bytes
802faba | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 | # UFUSC: Unified Federated Unlearning via Sensitivity-Guided Contrastive Forgetting
**A Novel Research Contribution in Machine Unlearning**
## Paper
📄 **Title:** Sensitivity-Guided Contrastive Forgetting: Unified Label and Feature Unlearning in Vertical Federated Learning
This repository contains the complete research artifacts for UFUSC — the **first framework to simultaneously perform label AND feature unlearning in Vertical Federated Learning (VFL)**.
## Key Innovation
Existing federated unlearning methods address either:
- **Label unlearning in VFL** (Manifold Mixup, arxiv:2410.10922) — forgets class labels but not features
- **Feature unlearning in HFL** (Ferrari, arxiv:2405.17462) — forgets features but in horizontal FL only
**UFUSC unifies both** through three innovations:
1. 🎯 **Contrastive Forgetting Loss (CFL)** — Repels forget-set embeddings from class centroids while anchoring retain-set representations
2. 📐 **Lipschitz Feature Sensitivity** — Minimizes model responsiveness to target features via perturbation-based sensitivity
3. ✅ **Dual-Variable Certification** — Provides convergence-based forgetting guarantees via primal-dual optimization
## Results Summary
| Dataset | Method | Retain Acc ↑ | Forget Acc ↓ | MIA ASR ↓ |
|---|---|---|---|---|
| MNIST | UFUSC-Joint | 85.45% | **13.00%** | **29.70%** |
| F-MNIST | UFUSC-Joint | 70.18% | **3.00%** | **19.10%** |
| CIFAR-10 | UFUSC-Joint | 50.96% | **4.80%** | **38.80%** |
UFUSC-Joint achieves the **lowest forget accuracy and MIA attack success rate** across all baselines.
## Repository Contents
- `paper.md` — Full conference-ready research paper (NeurIPS/ICML style)
- `research_paper.py` — Complete self-contained implementation (baselines + UFUSC + experiments + visualization)
- `results/` — All experimental results in JSON format
- `figures/` — Publication-quality visualizations
## Quick Start
```bash
pip install torch torchvision numpy matplotlib seaborn pandas scikit-learn
python research_paper.py
```
## Citation
If you use this work, please cite both anchor papers:
```bibtex
@article{bryan2024vfl_label_unlearning,
title={Towards Privacy-Guaranteed Label Unlearning in Vertical Federated Learning},
author={Bryan, H.X. et al.},
journal={arXiv preprint arXiv:2410.10922},
year={2024}
}
@article{ong2024ferrari,
title={Ferrari: Federated Feature Unlearning via Optimizing Feature Sensitivity},
author={Ong, W.K. et al.},
journal={arXiv preprint arXiv:2405.17462},
year={2024}
}
```
## License
This research is released for academic purposes. See paper for full details.
|