Research map / Federated learning and privacy / Personalized and heterogeneous FL
GDFA: Geometry-Driven Federated Unlearning with Directional Task Vector Alignment
Xiuting Weng, Ruizhi Pu, Yuanhang Yao, Kun Yue, Zhiwen Tang, Lixing Yu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| negate the merged task vector with larger magnitude (lambda from 0 to -1) (ablation) | GDFA · FMNIST (IID) | FA: 98.21 -> 1.45 | Improved |
| migrate global model to flat region via perturbation radius rho = 0.5 (ablation) | GDFA · CIFAR10 (Dir 1) | RA / FA | Improved |
| use too small perturbation radius rho (ablation) | GDFA · CIFAR10 (Dir 1) | RA / FA | Hurt |
| use too large perturbation radius rho (ablation) | GDFA · CIFAR10 (Dir 1) | RA | Hurt |
| increase number of perturbed models K (use K = 5) (ablation) | GDFA · CIFAR10 (Dir 1) | RA / time | Mixed |
| flat-region migration plus sign-consensus task vectors vs SIFU on retained accuracy | GDFA · CIFAR10 (Dir 0.5) | RA: 55.17 -> 69.33 | Improved |
| GDFA versus retraining on unlearning time | GDFA · CIFAR10 (Dir 0.5) | time (s): 1161.33 -> 31.92 | Improved |
| GDFA versus Eraser on forgotten accuracy | GDFA · CIFAR10 (Dir 0.5) | FA: 0.68 -> 1.57 | Hurt |
| GDFA versus Retrain on IID CIFAR10 forgotten accuracy | GDFA · CIFAR10 (IID) | FA: 2.32 -> 3.04 | Hurt |
Closest papers by meaning
- Unlearning through Knowledge Overwriting: Reversible Federated Unlearning via Selective Sparse Adapter (CVPR 2025 · similarity 0.89)
- Federated Unlearning with Gradient Descent and Conflict Mitigation (AAAI 2025 · similarity 0.89)
- FLUX: Efficient Descriptor-Driven Clustered Federated Learning under Arbitrary Distribution Shifts (NeurIPS 2025 · similarity 0.87)
- NoT: Federated Unlearning via Weight Negation (CVPR 2025 · similarity 0.87)
- Ferrari: Federated Feature Unlearning via Optimizing Feature Sensitivity (NeurIPS 2024 · similarity 0.86)
- Overcoming Data and Model heterogeneities in Decentralized Federated Learning via Synthetic Anchors (ICML 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
