Distributed optimization with faulty nodes: robust aggregation in hyperbolic space

The increasing deployment of distributed machine learning models necessitates robust optimization methods that can tolerate adversarial or faulty nodes. In this work, we propose a robust gradient aggregation method for distributed stochastic gradient descent that leverages hyperbolic geometry. Speci...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Neural computing & applications Jg. 37; H. 26; S. 21563 - 21605
Hauptverfasser: Ghosh, Subhas Kumar, Vittamsetti, Vijay Monic
Format: Journal Article
Sprache:Englisch
Veröffentlicht: London Springer London 01.09.2025
Springer Nature B.V
Schlagworte:
Accuracy
Artificial Intelligence
Communication
Computation
Computational Biology/Bioinformatics
Computational Science and Engineering
Computer Science
Data Mining and Knowledge Discovery
Euclidean geometry
Euclidean space
Federated learning
Hyperbolic coordinates
Image Processing and Computer Vision
Machine learning
Network management systems
Nodes
Optimization
Optimization algorithms
Original Article
Privacy
Probability and Statistics in Computer Science
Robustness
Sensors
Traffic flow
Distributed optimization algorithm
Federated learning
Stochastic gradient descent
Robust aggregation rule
Byzantine resilience
Hyperbolic space
ISSN:0941-0643, 1433-3058
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:The increasing deployment of distributed machine learning models necessitates robust optimization methods that can tolerate adversarial or faulty nodes. In this work, we propose a robust gradient aggregation method for distributed stochastic gradient descent that leverages hyperbolic geometry. Specifically, local gradients computed at individual nodes are embedded into hyperbolic space using the Poincaré ball model, and their geometric median is computed as a robust aggregate. This aggregated gradient is then mapped back to Euclidean space for the gradient update. We also show that existing robust gradient aggregation methods like Krum can be improved using hyperbolic space. Compared to existing robust aggregation methods, our hyperbolic approach offers improved separation of outlier updates. We provide theoretical convergence guarantees and validate our method on benchmark datasets as well as on a traffic forecasting task, demonstrating its efficacy in mitigating Byzantine failures in distributed federated learning environments.
Bibliographie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:0941-0643
1433-3058
DOI:10.1007/s00521-025-11475-0