Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Recurso digital |
| Language: | |
| Published: |
Zenodo
2025
|
| Online Access: | https://doi.org/10.5281/zenodo.17926595 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866901677321224192 |
|---|---|
| author | Ruksar Fatima Aliza Mahvash Ayesha Siddiqua Syeda Sheeba |
| author_facet | Ruksar Fatima Aliza Mahvash Ayesha Siddiqua Syeda Sheeba |
| contents | <p>Federated Learning (FL) has emerged as a powerful paradigm that enables collaborative model training across decentralized clients while preserving data privacy. Instead of aggregating sensitive data in a central server, FL coordinates local training on distributed devices—ranging from smartphones to IoT sensors and institutional servers—and collects only model updates. This design addresses major privacy, ethical, and security concerns associated with centralized data storage. Between 2019 and 2024, extensive research has focused on core FL challenges such as non-IID data distributions, device and system heterogeneity, resource limitations, privacy risks arising from gradient leakage, and practical deployment barriers in fields like healthcare and edge IoT. In this paper, we review recent advances across four themes: (1) distinctions and best practices for cross-device vs. cross-silo FL; (2) privacy-preserving mechanisms, including differential privacy and secure aggregation; (3) communication- and model-compression techniques for reducing bandwidth usage; and (4) real-world deployments in healthcare and edge-IoT environments. We analyze these works based on efficiency, accuracy, privacy trade-offs, and deployment-level considerations such as resource savings and regulatory alignment. Our synthesis shows that modern compression techniques—such as quantization, sparsification, and knowledge distillation—can significantly reduce communication costs with minimal accuracy loss, making FL feasible for resource-constrained devices. Privacy mechanisms remain essential for sensitive domains, though they commonly introduce accuracy and utility trade-offs. Cross-silo deployments demonstrate performance close to centralized baselines while maintaining data locality, yet full-scale adoption still depends on standardization, infrastructure readiness, and clearer ROI evidence. We highlight open gaps such as limited convergence theory for private and compressed FL under non-IID data, lack of unified benchmarks, insufficient empirical ROI studies, and the challenges of scaling FL to large modern models. To support clarity, we provide comparative tables, research-gap matrices, and conceptual diagrams illustrating accuracy-efficiency-privacy tradeoffs, followed by prioritized directions for future research.</p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_17926595 |
| institution | Zenodo |
| language | |
| publishDate | 2025 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | Federated Learning: Advances, Privacy Mechanisms, and Real-World Deployments Ruksar Fatima Aliza Mahvash Ayesha Siddiqua Syeda Sheeba <p>Federated Learning (FL) has emerged as a powerful paradigm that enables collaborative model training across decentralized clients while preserving data privacy. Instead of aggregating sensitive data in a central server, FL coordinates local training on distributed devices—ranging from smartphones to IoT sensors and institutional servers—and collects only model updates. This design addresses major privacy, ethical, and security concerns associated with centralized data storage. Between 2019 and 2024, extensive research has focused on core FL challenges such as non-IID data distributions, device and system heterogeneity, resource limitations, privacy risks arising from gradient leakage, and practical deployment barriers in fields like healthcare and edge IoT. In this paper, we review recent advances across four themes: (1) distinctions and best practices for cross-device vs. cross-silo FL; (2) privacy-preserving mechanisms, including differential privacy and secure aggregation; (3) communication- and model-compression techniques for reducing bandwidth usage; and (4) real-world deployments in healthcare and edge-IoT environments. We analyze these works based on efficiency, accuracy, privacy trade-offs, and deployment-level considerations such as resource savings and regulatory alignment. Our synthesis shows that modern compression techniques—such as quantization, sparsification, and knowledge distillation—can significantly reduce communication costs with minimal accuracy loss, making FL feasible for resource-constrained devices. Privacy mechanisms remain essential for sensitive domains, though they commonly introduce accuracy and utility trade-offs. Cross-silo deployments demonstrate performance close to centralized baselines while maintaining data locality, yet full-scale adoption still depends on standardization, infrastructure readiness, and clearer ROI evidence. We highlight open gaps such as limited convergence theory for private and compressed FL under non-IID data, lack of unified benchmarks, insufficient empirical ROI studies, and the challenges of scaling FL to large modern models. To support clarity, we provide comparative tables, research-gap matrices, and conceptual diagrams illustrating accuracy-efficiency-privacy tradeoffs, followed by prioritized directions for future research.</p> |
| title | Federated Learning: Advances, Privacy Mechanisms, and Real-World Deployments |
| url | https://doi.org/10.5281/zenodo.17926595 |