arXiv:2609.21057v1 Announce Type: cross Abstract: Federated learning (FL) enables collaborative model training without sharing raw data, but its performance degrades under non-IID data and stochastic client participation. Remedies built on classical Federated Averaging (FedAvg) typically presuppose that client participation probabilities are known to the server, which is rarely the case in deploy
FedeRage (Federated Risk-Averse Averaging) is a federated learning algorithm introduced by Herlock Rahimi and Dionysis Kalogerias in September 2026 that ensures provable convergence under unknown, skewed client participation and non-IID data. It addresses performance degradation caused by general client drift by embedding Conditional Value-at-Risk (CVaR) into the local objective within a distributionally robust optimization framework.
The method implicitly upweights high-loss and infrequently participating clients by transmitting only a single scalar per-client (an estimate of a loss quantile) alongside model parameters. This approach achieves an $\mathcal{O}(\kappa/\sqrt{T})$ convergence rate, where $\kappa$ represents the "price" of risk aversion, and remains agnostic to the participation distribution unlike optimal transport-based methods.
Experiments on MNIST, FashionMNIST, and CIFAR-10 benchmarks demonstrate that FedeRage improves upon state-of-the-art methods (such as FedAvg, FedProx, and SCAFFOLD) in terms of accuracy, fairness, and convergence speed, particularly in regimes with restricted and imbalanced client availability.
This paper addresses a central practical weakness of federated learning: standard Federated Averaging-style methods often assume either sufficiently homogeneous client data or known server-side control over client participation. In real deployments, both assumptions are fragile. Client data distributions can be non-IID and may drift over time, while participation is stochastic because devices come and go, experience connectivity issues, or are selected randomly. The authors argue that many existing fixes for FedAvg still rely on the server knowing the client participation probabilities, an assumption that is rarely available in operational systems.
The proposed method, FedeRage, is positioned as an agnostic federated learning framework designed to remain provably convergent under general client drift and uncertain participation. Rather than requiring detailed knowledge of the participation mechanism or imposing strong stationarity assumptions on the client population, the paper develops convergence theory for a broader setting in which client heterogeneity and selection uncertainty can vary in more general ways. The key insight is that robust convergence can be established without the server needing to explicitly model or estimate the underlying client participation law, which makes the approach more aligned with the realities of large-scale, dynamic federated systems.
The work matters because it moves federated learning theory closer to deployment conditions where data is heterogeneous, nonstationary, and client availability is unpredictable. If the claimed guarantees hold in practical regimes, FedeRage would provide a more trustworthy foundation for federated training on mobile devices, sensors, or other edge systems where clients join and leave unpredictably and local data distributions shift over time. Its contribution is therefore both methodological and conceptual: it reduces reliance on idealized participation assumptions and offers a path toward federated optimization guarantees that are more robust to the statistical drift inherent in real-world client populations.