Research map / Federated learning and privacy / Personalized and heterogeneous FL
REMISVFU: Vertical Federated Unlearning via Representation Misdirection for Intermediate Output Feature
Wenhan Wu, Zhili He, Huanghuang Liang, Yili Gong, Jiawei Jiang, Chuang Hu, Dazhao Cheng
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 |
|---|---|---|---|
| add orthogonal gradient projection between forgetting and retention gradients (ablation) | ReMisVFU · SVHN | clean accuracy: 65.04 -> 67.56 | Improved |
| add orthogonal gradient projection between forgetting and retention gradients (ablation) | ReMisVFU · SVHN | backdoor success rate (lower better): 11.65 -> 10.22 | Improved |
| add orthogonal gradient projection between forgetting and retention gradients (ablation) | ReMisVFU · CIFAR-10 | clean accuracy: 57.54 -> 58.45 | Improved |
| add orthogonal gradient projection between forgetting and retention gradients (ablation) | ReMisVFU · CIFAR-10 | backdoor success rate (lower better): 14.07 -> 10.55 | Improved |
| use orthogonal projection instead of a random unit vector of same norm (ablation) | ReMisVFU · SVHN | clean accuracy: 35.62 -> 67.56 | Improved |
| use orthogonal projection instead of a random unit vector of same norm (ablation) | ReMisVFU · CIFAR-10 | backdoor success rate (lower better): 87.21 -> 10.55 | Improved |
| increase anchor scale c to 4 or more (ablation) | ReMisVFU · SVHN and CIFAR-10 | forgetting (backdoor success) | Hurt |
| increase anchor scale c up to 2 (from 0.5) (ablation) | ReMisVFU · SVHN and CIFAR-10 | clean accuracy and backdoor success | No clear effect |
| increase retention weight alpha above 5x10^-3 (ablation) | ReMisVFU · SVHN and CIFAR-10 | backdoor success rate: exceeds 50% | Hurt |
| increase alpha from 1x10^-4 to 1x10^-3 (ablation) | ReMisVFU · SVHN and CIFAR-10 | clean accuracy | Improved |
| use ReMisVFU instead of FedGA gradient-ascent unlearning | FedGA · five image benchmarks | membership inference AUC: 1.000 -> 0.576 (CIFAR-10) | Improved |
| use ReMisVFU instead of full retraining | FedRetrain · five image benchmarks | unlearning time: 10.63%-21.49% of FedRetrain time | Improved |
Closest papers by meaning
- Federated Unlearning with Gradient Descent and Conflict Mitigation (AAAI 2025 · similarity 0.90)
- NoT: Federated Unlearning via Weight Negation (CVPR 2025 · similarity 0.88)
- Unlearning through Knowledge Overwriting: Reversible Federated Unlearning via Selective Sparse Adapter (CVPR 2025 · similarity 0.88)
- VFLAIR: A Research Library and Benchmark for Vertical Federated Learning (ICLR 2024 · similarity 0.86)
- Ferrari: Federated Feature Unlearning via Optimizing Feature Sensitivity (NeurIPS 2024 · similarity 0.86)
- Recurrent Early Exits for Federated Learning with Heterogeneous Clients (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.
