We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.NE

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Machine Learning

Title: Robust Federated Learning by Mixture of Experts

Abstract: We present a novel weighted average model based on the mixture of experts (MoE) concept to provide robustness in Federated learning (FL) against the poisoned/corrupted/outdated local models. These threats along with the non-IID nature of data sets can considerably diminish the accuracy of the FL model. Our proposed MoE-FL setup relies on the trust between users and the server where the users share a portion of their public data sets with the server. The server applies a robust aggregation method by solving the optimization problem or the Softmax method to highlight the outlier cases and to reduce their adverse effect on the FL process. Our experiments illustrate that MoE-FL outperforms the performance of the traditional aggregation approach for high rate of poisoned data from attackers.
Subjects: Machine Learning (cs.LG); Distributed, Parallel, and Cluster Computing (cs.DC); Neural and Evolutionary Computing (cs.NE); Networking and Internet Architecture (cs.NI)
Cite as: arXiv:2104.11700 [cs.LG]
  (or arXiv:2104.11700v1 [cs.LG] for this version)

Submission history

From: Saeedeh Parsaeefard [view email]
[v1] Fri, 23 Apr 2021 16:41:04 GMT (2229kb,D)

Link back to: arXiv, form interface, contact.