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

Momentum-Driven Adaptivity: Towards Tuning-Free Asynchronous Federated Learning

ICML 2025 · Federated learning and privacy › Communication-efficient FL · approach: Client participation analysis

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 MasFL and AdaMasFL, momentum-driven asynchronous federated learning methods that drop bounded data heterogeneity assumptions and, in the adaptive variant, remove all problem-specific learning-rate tuning via gradient normalization.

Problem

Asynchronous federated learning suffers from client drift under arbitrary data heterogeneity and requires learning rates tuned using unknown problem parameters (smoothness, heterogeneity bounds, delays).

Approach

A two-level (server and client) momentum with control variates handles staleness and heterogeneity without bounded-dissimilarity assumptions; AdaMasFL adds normalized local gradient steps so all stepsizes are set from S, K, and T alone.

Key results

• MasFL achieves O(1/sqrt(SKT)) convergence rate with linear speedup in clients and local steps, without bounded gradient dissimilarity assumptions.
• AdaMasFL achieves O(1/(SKT)^(1/4)) rate with all stepsizes determined by S, K, T alone, removing problem-parameter tuning.
• On non-i.i.d. CIFAR-10 and FMNIST, MasFL and AdaMasFL outperform CA2FL and FADAS in convergence speed and test accuracy, and AdaMasFL stays robust across learning rates and concurrency levels Mc.

Measured effects reported in the paper

Limitations

• MasFL still requires tuning of learning rates depending on unknown L and tau_max.
• Theoretical guarantees for AdaMasFL rely on a sqrt(SK T)-type scaling with slower rate O(1/(SKT)^(1/4)) than MasFL's O(1/sqrt(SKT)) in the gradient-norm metric.
• Experiments limited to image classification on CIFAR-10 and FMNIST with simulated asynchrony.

Experimental setup

Datasets: CIFAR-10 FMNIST

Benchmarks: CIFAR-10 FMNIST

Metrics: test accuracy convergence rate of average gradient norm

Baselines: CA2FL FADAS

Builds on: FedBuff SCAFFOLD CA2FL FADAS PAdaMFed

Closest papers by meaning

• Efficient Asynchronous Federated Learning with Prospective Momentum Aggregation and Fine-Grained Correction (AAAI 2024 · similarity 0.93)
• Momentum Benefits Non-iid Federated Learning Simply and Provably (ICLR 2024 · similarity 0.90)
• On the Role of Server Momentum in Federated Learning (AAAI 2024 · similarity 0.90)
• FedWMSAM: Fast and Flat Federated Learning via Weighted Momentum and Sharpness-Aware Minimization (NeurIPS 2025 · similarity 0.90)
• FedASMU: Efficient Asynchronous Federated Learning with Dynamic Staleness-Aware Model Update (AAAI 2024 · similarity 0.90)
• Efficient Federated Learning against Heterogeneous and Non-stationary Client Unavailability (NeurIPS 2024 · similarity 0.89)