README.md

# 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.