Federated Long-Tailed Learning by Retraining the Biased Classifier with Prototypes

Yang Li; Kan Li

doi:10.1007/978-981-99-9331-4_38

Federated Long-Tailed Learning by Retraining the Biased Classifier with Prototypes

Yang Li, Kan Li^*

^*Corresponding author for this work

School of Computer Science and Technology

Beijing Institute of Technology

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

1 Citation (Scopus)

Abstract

Federated learning is a privacy-preserving framework that collaboratively trains the global model without sharing raw data among clients. However, one significant issue encountered in federated learning is that biased classifiers affect the classification performance of the global model, especially when training on long-tailed data. Retraining the classifier on balanced datasets requires sharing the client’s information and poses the risk of privacy leakage. We propose a method for retraining the biased classifier using prototypes, that leverage the comparison of distances between local and global prototypes to guide the local training process. We conduct experiments on CIFAR-10-LT and CIFAR-100-LT, and our approach outperforms the accuracy of baseline methods, with accuracy improvements of up to 10%.

Original language	English
Title of host publication	Frontiers in Cyber Security - 6th International Conference, FCS 2023, Revised Selected Papers
Editors	Haomiao Yang, Rongxing Lu
Publisher	Springer Science and Business Media Deutschland GmbH
Pages	575-585
Number of pages	11
ISBN (Print)	9789819993307
DOIs	https://doi.org/10.1007/978-981-99-9331-4_38
Publication status	Published - 2024
Event	6th International Conference on Frontiers in Cyber Security, FCS 2023 - Chengdu, China Duration: 21 Aug 2023 → 23 Aug 2023

Publication series

Name	Communications in Computer and Information Science
Volume	1992
ISSN (Print)	1865-0929
ISSN (Electronic)	1865-0937

Conference

Conference	6th International Conference on Frontiers in Cyber Security, FCS 2023
Country/Territory	China
City	Chengdu
Period	21/08/23 → 23/08/23

Keywords

Federated learning
Long-tailed data
Privacy protection
Prototype learning

Access to Document

10.1007/978-981-99-9331-4_38

Cite this

Li, Y., & Li, K. (2024). Federated Long-Tailed Learning by Retraining the Biased Classifier with Prototypes. In H. Yang, & R. Lu (Eds.), Frontiers in Cyber Security - 6th International Conference, FCS 2023, Revised Selected Papers (pp. 575-585). (Communications in Computer and Information Science; Vol. 1992). Springer Science and Business Media Deutschland GmbH. https://doi.org/10.1007/978-981-99-9331-4_38

@inproceedings{f56ffe8c430f40dc86611aa727ada3de,

title = "Federated Long-Tailed Learning by Retraining the Biased Classifier with Prototypes",

abstract = "Federated learning is a privacy-preserving framework that collaboratively trains the global model without sharing raw data among clients. However, one significant issue encountered in federated learning is that biased classifiers affect the classification performance of the global model, especially when training on long-tailed data. Retraining the classifier on balanced datasets requires sharing the client{\textquoteright}s information and poses the risk of privacy leakage. We propose a method for retraining the biased classifier using prototypes, that leverage the comparison of distances between local and global prototypes to guide the local training process. We conduct experiments on CIFAR-10-LT and CIFAR-100-LT, and our approach outperforms the accuracy of baseline methods, with accuracy improvements of up to 10%.",

keywords = "Federated learning, Long-tailed data, Privacy protection, Prototype learning",

author = "Yang Li and Kan Li",

note = "Publisher Copyright: {\textcopyright} The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd 2024.; 6th International Conference on Frontiers in Cyber Security, FCS 2023 ; Conference date: 21-08-2023 Through 23-08-2023",

year = "2024",

doi = "10.1007/978-981-99-9331-4_38",

language = "English",

isbn = "9789819993307",

series = "Communications in Computer and Information Science",

publisher = "Springer Science and Business Media Deutschland GmbH",

pages = "575--585",

editor = "Haomiao Yang and Rongxing Lu",

booktitle = "Frontiers in Cyber Security - 6th International Conference, FCS 2023, Revised Selected Papers",

address = "Germany",

}

Li, Y & Li, K 2024, Federated Long-Tailed Learning by Retraining the Biased Classifier with Prototypes. in H Yang & R Lu (eds), Frontiers in Cyber Security - 6th International Conference, FCS 2023, Revised Selected Papers. Communications in Computer and Information Science, vol. 1992, Springer Science and Business Media Deutschland GmbH, pp. 575-585, 6th International Conference on Frontiers in Cyber Security, FCS 2023, Chengdu, China, 21/08/23. https://doi.org/10.1007/978-981-99-9331-4_38

Federated Long-Tailed Learning by Retraining the Biased Classifier with Prototypes. / Li, Yang; Li, Kan.
Frontiers in Cyber Security - 6th International Conference, FCS 2023, Revised Selected Papers. ed. / Haomiao Yang; Rongxing Lu. Springer Science and Business Media Deutschland GmbH, 2024. p. 575-585 (Communications in Computer and Information Science; Vol. 1992).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

TY - GEN

T1 - Federated Long-Tailed Learning by Retraining the Biased Classifier with Prototypes

AU - Li, Yang

AU - Li, Kan

N1 - Publisher Copyright: © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd 2024.

PY - 2024

Y1 - 2024

N2 - Federated learning is a privacy-preserving framework that collaboratively trains the global model without sharing raw data among clients. However, one significant issue encountered in federated learning is that biased classifiers affect the classification performance of the global model, especially when training on long-tailed data. Retraining the classifier on balanced datasets requires sharing the client’s information and poses the risk of privacy leakage. We propose a method for retraining the biased classifier using prototypes, that leverage the comparison of distances between local and global prototypes to guide the local training process. We conduct experiments on CIFAR-10-LT and CIFAR-100-LT, and our approach outperforms the accuracy of baseline methods, with accuracy improvements of up to 10%.

AB - Federated learning is a privacy-preserving framework that collaboratively trains the global model without sharing raw data among clients. However, one significant issue encountered in federated learning is that biased classifiers affect the classification performance of the global model, especially when training on long-tailed data. Retraining the classifier on balanced datasets requires sharing the client’s information and poses the risk of privacy leakage. We propose a method for retraining the biased classifier using prototypes, that leverage the comparison of distances between local and global prototypes to guide the local training process. We conduct experiments on CIFAR-10-LT and CIFAR-100-LT, and our approach outperforms the accuracy of baseline methods, with accuracy improvements of up to 10%.

KW - Federated learning

KW - Long-tailed data

KW - Privacy protection

KW - Prototype learning

UR - http://www.scopus.com/inward/record.url?scp=85181975846&partnerID=8YFLogxK

U2 - 10.1007/978-981-99-9331-4_38

DO - 10.1007/978-981-99-9331-4_38

M3 - Conference contribution

AN - SCOPUS:85181975846

SN - 9789819993307

T3 - Communications in Computer and Information Science

SP - 575

EP - 585

BT - Frontiers in Cyber Security - 6th International Conference, FCS 2023, Revised Selected Papers

A2 - Yang, Haomiao

A2 - Lu, Rongxing

PB - Springer Science and Business Media Deutschland GmbH

T2 - 6th International Conference on Frontiers in Cyber Security, FCS 2023

Y2 - 21 August 2023 through 23 August 2023

ER -

Li Y, Li K. Federated Long-Tailed Learning by Retraining the Biased Classifier with Prototypes. In Yang H, Lu R, editors, Frontiers in Cyber Security - 6th International Conference, FCS 2023, Revised Selected Papers. Springer Science and Business Media Deutschland GmbH. 2024. p. 575-585. (Communications in Computer and Information Science). doi: 10.1007/978-981-99-9331-4_38