Research map / Federated learning and privacy / Communication-efficient FL
A-FedPD: Aligning Dual-Drift is All Federated Primal-Dual Learning Needs
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 A-FedPD, a federated primal-dual method that virtually updates dual variables of inactive clients using the aggregated global model to counter "dual drift" under partial participation, with non-convex convergence and generalization analysis and CIFAR experiments.
Problem
Federated primal-dual methods suffer instability under partial participation because dual variables of long-inactive clients become stale and mismatched with the global model ("dual drift").
Approach
Construct virtual dual updates for unparticipated clients using the aggregated active-client model, so their dual variables stay aligned with the global consensus, approximating quasi-full participation.
Key results
- A-FedPD reaches 85.31% on CIFAR-10 IID with LeNet (SGD), vs 84.49% for FedDyn and 84.25% for SCAFFOLD; A-FedPDSAM reaches 86.47%.
- On CIFAR-100 (ResNet-18 setting, Dir-0.1), A-FedPDSAM reaches 53.15% while FedDyn becomes unstable (loss oscillates).
- A-FedPD needs 131 rounds to reach 81% on CIFAR-10 vs 501 for FedAvg (3.82x speedup).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace local SGD with SAM (ablation) | A-FedPD · CIFAR-10 and CIFAR-100, IID/Dir-1.0/Dir-0.1, LeNet and ResNet-18 | test accuracy: 85.31 -> 86.47; 84.94 -> 85.90; 80.28 -> 81.96; 51.41 -> 55.56; 51.17 -> 54.62; 48.15 -> 53.15; 85.11 -> 87.44; 84.33 -> | Helped |
| Use virtual dual updates for unparticipated clients | A-FedPD · CIFAR-10 and CIFAR-100 | communication rounds to target accuracy: 501 -> 131; 772 -> 123 | Helped |
| Use SAM instead of SGD for local optimization (ablation) | A-FedPD · CIFAR-10 and CIFAR-100, LeNet and ResNet-18 | seconds per round: 37.61 -> 60.63; 11.71 -> 13.90 | Hurt |
Surprising findings
- FedDyn's loss oscillates on ResNet-18 CIFAR-100 under its training setup, reported as a dual drift failure.
Limitations
- Server must store and update dual variables of all clients, increasing storage cost.
- Per-round computation is higher than FedAvg and SCAFFOLD, making early wall-clock performance weaker.
- Theory relies on smoothness and Lipschitz assumptions and an epsilon-inexact local solver.
Experimental setup
Datasets: CIFAR-10 CIFAR-100
Benchmarks: CIFAR-10 CIFAR-100
Metrics: test accuracy communication rounds to target accuracy wall-clock time
Baselines: FedAvg FedCM SCAFFOLD FedSAM FedDyn FedSpeed
Builds on: FedADMM FedDyn FedPD FedSpeed SAM
Closest papers by meaning
- A Primal-Dual Algorithm for Hybrid Federated Learning (AAAI 2024 · similarity 0.88)
- PubSub-VFL: Towards Efficient Two-Party Split Learning in Heterogeneous Environments via Publisher/Subscriber Architecture (NeurIPS 2025 · similarity 0.87)
- Efficient Federated Learning against Byzantine Attacks and Data Heterogeneity via Aggregating Normalized Gradients (NeurIPS 2025 · similarity 0.87)
- FedLMT: Tackling System Heterogeneity of Federated Learning via Low-Rank Model Training with Theoretical Guarantees (ICML 2024 · similarity 0.87)
- Exact and Linear Convergence for Federated Learning under Arbitrary Client Participation is Attainable (NeurIPS 2025 · similarity 0.87)
- Convergence Analysis of Split Federated Learning on Heterogeneous Data (NeurIPS 2024 · similarity 0.87)
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.