Research map / Federated learning and privacy / Communication-efficient FL
PDUDT: Provable Decentralized Unlearning under Dynamic Topologies
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PDUDT, a retraining-free decentralized unlearning method under dynamic topologies that corrects each client's model using cached historical gradient residuals plus Gaussian noise, with indistinguishability and O(1/T) convergence guarantees.
Problem
Removing a departing client's influence from a decentralized learning system whose communication topology changes over rounds, without retraining from scratch or trusting a central server.
Approach
Each client subtracts a weighted sum of cached gradient-residual approximations (weights proportional to squared norms of neighbor-gradient aggregates) from its model at the unlearning round, then adds Gaussian noise so the result is (epsilon, beta)-indistinguishable from perturbed retraining; no extra communication or retraining is needed.
Key results
- PDUDT reduces unlearning time by over 99% versus Retrain (e.g., 3.0 s vs 1322.2 s on MNIST) while achieving comparable accuracy to perturbed retraining across noise scales.
- After unlearning, membership inference attack precision drops to about 49-51% for PDUDT, close to Retrain (about 49-51%), versus 65-68% for the original model.
- Theory shows subsequent decentralized learning converges at rate O(1/T), claimed to match state-of-the-art results.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PDUDT instead of Retrain | Decentralized unlearning · MNIST | unlearning time (s): 1322.2 -> 3.0 | Helped |
| Use PDUDT instead of Retrain | Decentralized unlearning · Fashion-MNIST | unlearning time (s): 1299.8 -> 3.3 | Helped |
| Use PDUDT instead of Retrain | Decentralized unlearning · CIFAR-10 | unlearning time (s): 1696.2 -> 9.6 | Helped |
| Use PDUDT instead of Retrain | Decentralized unlearning · SVHN | unlearning time (s): 2747.8 -> 9.8 | Helped |
| Add early stopping (ablation) | PDUDT · MNIST | unlearning time (s): 3.0 -> 2.3 | Helped |
| Add early stopping (ablation) | PDUDT · Fashion-MNIST | unlearning time (s): 3.3 -> 2.6 | Helped |
| Add early stopping (ablation) | PDUDT · CIFAR-10 | unlearning time (s): 9.6 -> 7.5 | Helped |
| Add early stopping (ablation) | PDUDT · SVHN | unlearning time (s): 9.8 -> 7.8 | Helped |
| Unlearn with PDUDT rather than leave the original model unchanged | CNN · MNIST | attack precision (%): 65.3 ± 0.9 -> 49.9 ± 0.6 | Helped |
| Unlearn with PDUDT rather than leave the original model unchanged | CNN · Fashion-MNIST | attack precision (%): 67.1 ± 1.9 -> 49.0 ± 1.3 | Helped |
| Unlearn with PDUDT rather than leave the original model unchanged | ResNet-18 · CIFAR-10 | attack precision (%): 65.3 ± 0.7 -> 50.1 ± 1.6 | Helped |
| Unlearn with PDUDT rather than leave the original model unchanged | ResNet-18 · SVHN | attack precision (%): 68.1 ± 0.5 -> 51.3 ± 0.3 | Helped |
Surprising findings
- PDUDT (ES), which keeps neighbor gradients only for the first 80 of 100 rounds, is slightly degraded on MNIST but otherwise comparable.
Limitations
- Space overhead grows linearly with rounds and neighbor count, requiring caching of neighbors' gradients.
- Experiments use only 10 clients and small image datasets with CNN and ResNet-18 models, starting from pre-trained models to control noise.
- Assumptions include Lipschitz gradients, bounded variance, and spectral gap conditions on doubly stochastic matrices.
Experimental setup
Datasets: MNIST Fashion-MNIST CIFAR-10 SVHN
Benchmarks: MNIST Fashion-MNIST CIFAR-10 SVHN
Metrics: accuracy per-class accuracy unlearning time membership inference attack precision
Baselines: Origin Retrain FATS-Unl FedRecovery HDUS
Builds on: decentralized SGD (Lian et al., 2017) FATS-Unl (Tao et al., 2024) FedRecovery (Zhang et al., 2023) Gaussian mechanism (Bun & Steinke, 2016) (epsilon, beta)-indistinguishability (Neel et al., 2021)
Closest papers by meaning
- Trajectory-Aware Certified Decentralized Unlearning via SGD Stability (ICML 2026 · similarity 0.89)
- DePRL: Achieving Linear Convergence Speedup in Personalized Decentralized Learning with Shared Representations (AAAI 2024 · similarity 0.86)
- Unextractable Protocol Models: Collaborative Training and Inference without Weight Materialization (NeurIPS 2025 · similarity 0.84)
- Unlearning through Knowledge Overwriting: Reversible Federated Unlearning via Selective Sparse Adapter (CVPR 2025 · similarity 0.84)
- Unveiling the Power of Multiple Gossip Steps: A Stability-Based Generalization Analysis in Decentralized Training (NeurIPS 2025 · similarity 0.84)
- NoT: Federated Unlearning via Weight Negation (CVPR 2025 · similarity 0.84)
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.