AI Paper Atlas

Research map / Federated learning and privacy / Communication-efficient FL

Enhancing Communication Compression via Discrepancy-aware Calibration for Federated Learning

Zhiyi Wan, Yijia Chi, Liang Li, Wanrou Du, Miao Pan, Xiaoqi Qin

ICLR 2026 · Federated learning and privacy › Communication-efficient FL · approach: Efficient local training in FL

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 changedWhereMeasuredEffect
replace magnitude-based selection with discrepancy-aware selectionTop-k · CIFAR-10 (ViT-tiny), ratio 0.01final test accuracy: 21.03 -> 29.62Improved
replace magnitude-based selection with discrepancy-aware selectionTop-k · CIFAR-100 (ViT-small), ratio 0.01final test accuracy: 9.29 -> 13.29Improved
replace magnitude-based selection with discrepancy-aware selectionTop-k · CIFAR-100 (ResNet-18), ratio 0.01final test accuracy: 10.76 -> 15.58Improved
replace magnitude-based selection with discrepancy-aware selectionTop-k · Fashion-MNIST (AlexNet), ratio 0.01final test accuracy: 63.31 -> 67.55Improved
replace magnitude-based selection with discrepancy-aware selectionTop-k · CIFAR-10 (ViT-tiny), ratio 0.6final test accuracy: 38.57 -> 41.21Improved
replace magnitude-based selection with discrepancy-aware selectionATOMO · CIFAR-100 (ViT-base), rank 1final test accuracy: 17.08 -> 20.17Improved
replace magnitude-based selection with discrepancy-aware selectionATOMO · CIFAR-10 (ViT-tiny), rank 8final test accuracy: 42.44 -> 42.68No clear effect
replace magnitude-based selection with discrepancy-aware selectionATOMO · CIFAR-10 (ViT-small), rank 2final test accuracy: 37.01 -> 40.10Improved
add discrepancy-aware selection to reduce rounds to 80% targetTop-k · CIFAR-10 (ViT-tiny), ratio 0.01communication rounds: 128 -> 82 (1.56x)Improved
add discrepancy-aware selection to reduce rounds to 80% targetTop-k · FMNIST (AlexNet), ratio 0.2communication rounds: 83 -> 83 (1.00x)No clear effect
use random calibration samples instead of first samples (ablation)discrepancy-aware Top-k · Fashion-MNIST (AlexNet)final test accuracy: 72.49 -> 73.95Improved
use random calibration samples instead of last samples (ablation)discrepancy-aware Top-k · CIFAR-10 (ViT-tiny)final test accuracy: 42.91 -> 43.23Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy