Research map / Federated learning and privacy / Communication-efficient FL
TurboSVM-FL: Boosting Federated Learning through SVM Aggregation for Lazy Clients
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 TurboSVM-FL, a server-side federated aggregation strategy that fits SVMs over client class embeddings, averages support-vector embeddings, and applies max-margin spread-out regularization, to speed convergence for lazy clients with no added client cost.
Problem
Slow convergence of federated learning under non-iid data when clients are computation-limited and train only a few local epochs, where client-side regularizers or extra computation are impractical.
Approach
Treats each client's class embeddings (rows of the logit layer) as samples for a higher-level one-vs-one SVM; aggregates only support-vector class embeddings, then applies a Gaussian-kernel spread-out regularizer on embeddings projected onto SVM hyperplanes, with the server optimizer Adam. All computation is on the server, so client cost is unchanged.
Key results
- TurboSVM-FL reaches 70% accuracy on FEMNIST in 54.6 rounds vs 144.4 for FedAvg (62.2% reduction), and 46.4 vs 91.6 rounds on CelebA (49.3% reduction).
- After 100 rounds on FEMNIST with C=8, E=1, TurboSVM-FL attains 76.6% accuracy, 61.9 F1 and 75.8 MCC, above all six baselines.
- On Shakespeare TurboSVM-FL reaches 50% accuracy in 43.4 rounds vs 51.0 for FedAvg (14.9% reduction), but its final accuracy (53.7%) is comparable to FedAdam and FedAMS (53.8%, 54.0%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TurboSVM-FL aggregation instead of FedAwS | Federated learning · FEMNIST | communication rounds to 70% accuracy: 81.4$\pm$2.2 -> 54.6$\pm$1.6 | Helped |
| Use TurboSVM-FL aggregation instead of FedAwS | Federated learning · CelebA | communication rounds to 70% accuracy: 84.2$\pm$24.4 -> 46.4$\pm$9.4 | Helped |
| Use TurboSVM-FL aggregation instead of FedAwS | Federated learning · Shakespeare | communication rounds to 50% accuracy: 45.0$\pm$2.3 -> 43.4$\pm$2.9 | Helped |
| Use TurboSVM-FL aggregation instead of FedAwS | Federated learning · FEMNIST | F1: 54.8$\pm$0.8 -> 61.9$\pm$0.7 | Helped |
| Use TurboSVM-FL aggregation instead of FedAwS | Federated learning · CelebA | F1: 72.2$\pm$3.0 -> 77.2$\pm$1.3 | Helped |
| Use TurboSVM-FL aggregation instead of FedAwS | Federated learning · Shakespeare | F1: 18.3$\pm$0.6 -> 19.2$\pm$0.3 | Helped |
| Use TurboSVM-FL aggregation instead of FedAMS | Federated learning · Shakespeare | accuracy: 54.0$\pm$0.5 -> 53.7$\pm$0.2 | Hurt |
| Increase embedding dimension (ablation) | TurboSVM-FL · FEMNIST, C=8 | F1: 6.1 -> 67.4 | Helped |
| Increase embedding dimension (ablation) | TurboSVM-FL · FEMNIST, C=64 | F1: 16.3 -> 68.3 | Helped |
| Increase embedding dimension (ablation) | TurboSVM-FL · FEMNIST, C=512 | F1: 16.7 -> 67.0 | Helped |
| Use polynomial kernel degree 4 instead of linear kernel (ablation) | TurboSVM-FL · CelebA | F1: 82.1 -> 85.2 | Helped |
| Use RBF kernel instead of linear kernel (ablation) | TurboSVM-FL · CelebA | F1: 82.1 -> 76.8 | Hurt |
Surprising findings
- On Shakespeare, FedAdam and FedAMS achieve equal or higher final accuracy and MCC than TurboSVM-FL, so the advantage is mainly in convergence speed rather than final performance on that task.
- FedAdam and FedAMS reach near-zero MCC on CelebA despite reasonable accuracy, indicating unstable or degenerate predictions.
Limitations
- Server must fit one-vs-one SVMs for every class pair (K(K-1)/2), which is computationally heavy for many classes; the authors suggest class sampling as mitigation.
- Linear separability assumed for class embeddings; kernelized variants did not yield consistent improvement in the appendix ablation.
- Theoretical analysis is informal and restricted to binary classification with simplifying assumptions (e.g., all clients are support vectors and equal dataset sizes).
Experimental setup
Datasets: FEMNIST CelebA Shakespeare
Benchmarks: FEMNIST CelebA Shakespeare
Metrics: communication rounds to target accuracy accuracy F1 score MCC
Baselines: FedAvg FedAdam FedAMS FedProx MOON FedAwS
Builds on: FedAvg FedOpt FedAwS FedProx MOON
Closest papers by meaning
- FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning (NeurIPS 2025 · similarity 0.90)
- FedAWA: Adaptive Optimization of Aggregation Weights in Federated Learning Using Client Vectors (CVPR 2025 · similarity 0.89)
- An Aggregation-Free Federated Learning for Tackling Data Heterogeneity (CVPR 2024 · similarity 0.89)
- FedSMU: Communication-Efficient and Generalization-Enhanced Federated Learning through Symbolic Model Updates (ICML 2025 · similarity 0.88)
- SparsyFed: Sparse Adaptive Federated Learning (ICLR 2025 · similarity 0.88)
- Towards the Theory of Unsupervised Federated Learning: Non-asymptotic Analysis of Federated EM Algorithms (ICML 2024 · similarity 0.88)
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.