Dynamic Regret for Byzantine-Robust Online Federated Learning
IEEE Transactions on Signal Processing · 2026 · 74 · 1357-1367
DOI: 10.1109/tsp.2026.3673260 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Signal Processing, 2026. Verfasst von Haibao Tian und 3 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 40 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Shanghai Municipal Science and Technology Major Project, Joint Funds of the National Natural Science Foundation of China.
Abstract (Originalsprache)
Federated learning enables decentralized model training across multiple clients without exchanging raw data, making it a crucial paradigm for privacy-preserving machine learning. However, its deployment in adversarial and dynamic environments remains fundamentally challenging, particularly under Byzantine attacks. Existing online Byzantine-robust methods face two critical limitations: (i) a strong reliance on the Independent and Identically Distributed (IID) assumption to achieve sublinear regret, and (ii) an inability to adapt to non-stationary environments due to their reliance on static regret metrics. These challenges substantially limit their applicability in practical federated learning scenarios characterized by heterogeneity and non-stationarity. To address these issues, this paper proposes a Byzantine-robust online algorithm, BR-OMGD, which performs multiple local gradient descent updates and incorporates a robust aggregation mechanism to mitigate adversarial effects under the weak growth condition. The algorithm achieves a near-optimal dynamic regret bound of $\mathcal{O}(S^{\ast}_{T})$ without requiring IID assumptions, where $S^{\ast}_{T}$ denotes the squared path-length. When the online problem reduces to the offline case, BR-OMGD further guarantees exact linear convergence under strong convexity, smoothness, and the weak growth condition, even in the presence of data heterogeneity and Byzantine adversaries. To the best of our knowledge, this is the first result establishing such convergence guarantees under these conditions. Extensive experiments on benchmark datasets (MNIST, CIFAR-100) demonstrate the proposed algorithm’s practical effectiveness, showing improved robustness and lower dynamic regret under Byzantine attacks.
Förderung
- National Natural Science Foundation of China
- Shanghai Municipal Science and Technology Major Project
- Joint Funds of the National Natural Science Foundation of China
Zitieren
APA
Tian, H., He, Q., Qu, Z., & Li, X. (2026). Dynamic Regret for Byzantine-Robust Online Federated Learning. IEEE Transactions on Signal Processing, 74, 1357-1367. https://doi.org/10.1109/tsp.2026.3673260
Vancouver
Tian H, He Q, Qu Z, Li X. Dynamic Regret for Byzantine-Robust Online Federated Learning. IEEE Transactions on Signal Processing. 2026;74:1357-1367. doi:10.1109/tsp.2026.3673260
BibTeX
@article{Tian2026,
author = {Tian, Haibao and He, Qiyong and Qu, Zhihai and Li, Xiuxian},
title = {Dynamic Regret for Byzantine-Robust Online Federated Learning},
journal = {IEEE Transactions on Signal Processing},
year = {2026},
volume = {74},
pages = {1357--1367},
doi = {10.1109/tsp.2026.3673260}
}
EUVIMED