Research map / Federated learning and privacy / Communication-efficient FL
SpecGradFilter: A Spectral Gradient Filtering Framework for Taming Federated Heterogeneity
Liyang Yuan, Yibo Yang, Dandan Guo, Peter Richtárik, Zhouchen Lin
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 |
|---|---|---|---|
| use low-frequency suppression instead of high-frequency suppression (ablation) | SpecGradFilter · CIFAR-10 (alpha=0.1) | accuracy: 59.99 -> 78.67 | Improved |
| use low-frequency suppression instead of random suppression (ablation) | SpecGradFilter · CIFAR-100 (alpha=0.1) | accuracy: 28.74 -> 40.87 | Improved |
| use spectral filtering instead of l2-norm rescaling (ablation) | SpecGradFilter · CIFAR-100 (alpha=0.1) | accuracy: 27.04 -> 40.87 | Improved |
| use FFT low-frequency suppression versus spatial Gaussian detrend (ablation) | SpecGradFilter · CIFAR-100 (alpha=0.1) | accuracy: 41.53 vs 40.87 | No clear effect |
| filter all convolutional layers instead of deep stage only (ablation) | SpecGradFilter · CIFAR-10 (alpha=0.1) | accuracy: 73.89 -> 78.67 | Improved |
| also filter classifier head (ablation) | SpecGradFilter · CIFAR-10 (alpha=0.1) | accuracy: 78.67 -> 78.58 | No clear effect |
| filter only patch embedding of ViT instead of all eligible weights (ablation) | SpecGradFilter · CIFAR-10 ViT (alpha=0.1) | accuracy: 44.77 -> 45.16 | Improved |
| filter attention in ViT (ablation) | SpecGradFilter · CIFAR-100 ViT (alpha=0.1) | accuracy: 23.11 -> 22.50 | Hurt |
| use natural flatten order 0123 instead of 3210 (ablation) | SpecGradFilter · CIFAR-10 (alpha=0.1) | accuracy: 55.87 -> 78.05 | Improved |
| use 1D FFT instead of 2D FFT (ablation) | SpecGradFilter · CIFAR-100 (alpha=0.1) | accuracy: 31.30 -> 40.16 | Improved |
| apply filter in centralized IID training | SpecGradFilter · CIFAR-100 centralized | accuracy: 68.66 -> 67.09 | Hurt |
| apply filter in near-IID federated learning alpha=100 | SpecGradFilter vs FedAvg · CIFAR-10 | accuracy: 81.28 -> 86.74 | Improved |
Closest papers by meaning
- FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning (NeurIPS 2025 · similarity 0.88)
- Federated Learning with Sample-level Client Drift Mitigation (AAAI 2025 · similarity 0.87)
- Locally Estimated Global Perturbations are Better than Local Perturbations for Federated Sharpness-aware Minimization (ICML 2024 · similarity 0.87)
- Decentralized Sporadic Federated Learning: A Unified Algorithmic Framework with Convergence Guarantees (ICLR 2025 · similarity 0.87)
- Federated Optimization with Doubly Regularized Drift Correction (ICML 2024 · similarity 0.87)
- FedAdaVR: Adaptive Variance Reduction for Robust Federated Learning under Limited Client Participation (NeurIPS 2026 · 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.
