Taxonomy and Resolution Strategy for Client-Level Disagreements in Federated Learning
A new paper on arXiv (2604.23386) introduces a taxonomy and resolution strategy for 'client-level disagreements' in Federated Learning (FL), where clients may need to exclude each other for strategic, regulatory, or competitive reasons. The authors propose a multi-track resolution strategy that creates isolated model update paths to prevent cross-contamination and unfairness. They validate their approach through simulations across 34 scenarios using MNIST and N-CMAPSS datasets, handling permanent, temporal, and overlapping disagreement patterns. The scalability analysis shows the server-side algorithm's efficiency.
Key facts
- Paper arXiv:2604.23386 addresses client-level disagreements in Federated Learning.
- Proposes a taxonomy of client-level disagreement scenarios.
- Introduces a multi-track resolution strategy with isolated model update paths.
- Strategy guarantees strict client exclusion and prevents cross-contamination.
- Empirical evaluation uses 34 scenarios with MNIST and N-CMAPSS datasets.
- Handles permanent, temporal, and overlapping disagreement patterns.
- Scalability analysis covers the server-side resolution algorithm.
- Published on arXiv with cross type announcement.
Entities
Institutions
- arXiv