Information: Join today and make your research activities more affordable! Technical workshop participation fees and annual registration fees are available at member rates.
Notice: [Important] Announcement of Changes to Registration Fee Payment and Manuscript Upload Procedures for IEICE Technical Meetings
IEICE Technical Committee Submission System
Conference Paper's Information		 Online Proceedings
[Sign in]
Tech. Rep. Archives
 Go Top Page Go Previous   [Japanese] / [English] 

Paper Abstract and Keywords
Presentation 2026-06-18 09:30
Coupled Adaptive Batch Sizes for One-Pass Local Differentially Private SGD
Bingchang He, Atsuko Miyaji (UOsaka) IA2026-14 ICSS2026-14
Abstract (in Japanese) (See Japanese page) 
(in English) Local differentially private stochastic gradient descent (LDP-SGD) is difficult to tune because each user typically contributes a privatized gradient only once, making repeated hyperparameter search costly in both privacy and sample efficiency. We study a one-pass ($varepsilon$, $delta$)-LDP-SGD scheme that adapts batch sizes online by coupling them to the learning rate and to estimates of the current objective value and privatized gradient variance. The method aims to reduce the need for repeated manual tuning of fixed batch sizes by using a single training run that allocates more users to updates where variance control is most valuable. We instantiate the approach in the LDP-SGD framework for linear models including logistic regression. Based on the variance reduction from gradient averaging in LDP and the robustness gains of coupled adaptive batch sizing, the proposed method outperforms most untuned fixed-batch configurations. These results would position adaptive batch sizing as a practical compromise between trainability and peak utility when each user can be used only once.
Keyword (in Japanese) (See Japanese page) 
(in English) local differential privacy / stochastic gradient descent / hyperparameter tuning / / / / /  
Reference Info. IEICE Tech. Rep., vol. 126, no. 74, ICSS2026-14, pp. 85-92, June 2026.
Paper # ICSS2026-14 
Date of Issue 2026-06-10 (IA, ICSS) 
ISSN Online edition: ISSN 2432-6380
Copyright
and
reproduction		 All rights are reserved and no part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission in writing from the publisher. Notwithstanding, instructors are permitted to photocopy isolated articles for noncommercial classroom use without fee. (License No.: 10GA0019/12GB0052/13GB0056/17GB0034/18GB0034)
Download PDF IA2026-14 ICSS2026-14

Conference Information
Committee IA ICSS  
Conference Date 2026-06-17 - 2026-06-18 
Place (in Japanese) (See Japanese page) 
Place (in English) e-Topia, Kagawa BB Square 
Topics (in Japanese) (See Japanese page) 
Topics (in English) Internet Security, etc. 
Paper Information
Registration To ICSS 
Conference Code 2026-06-IA-ICSS 
Language English 
Title (in Japanese) (See Japanese page) 
Sub Title (in Japanese) (See Japanese page) 
Title (in English) Coupled Adaptive Batch Sizes for One-Pass Local Differentially Private SGD 
Sub Title (in English)  
Keyword(1) local differential privacy  
Keyword(2) stochastic gradient descent  
Keyword(3) hyperparameter tuning  
Keyword(4)  
Keyword(5)  
Keyword(6)  
Keyword(7)  
Keyword(8)  
1st Author's Name Bingchang He  
1st Author's Affiliation The University of Osaka (UOsaka)
2nd Author's Name Atsuko Miyaji  
2nd Author's Affiliation The University of Osaka (UOsaka)
3rd Author's Name  
3rd Author's Affiliation ()
4th Author's Name  
4th Author's Affiliation ()
5th Author's Name  
5th Author's Affiliation ()
6th Author's Name  
6th Author's Affiliation ()
7th Author's Name  
7th Author's Affiliation ()
8th Author's Name  
8th Author's Affiliation ()
9th Author's Name  
9th Author's Affiliation ()
10th Author's Name  
10th Author's Affiliation ()
11th Author's Name  
11th Author's Affiliation ()
12th Author's Name  
12th Author's Affiliation ()
13th Author's Name  
13th Author's Affiliation ()
14th Author's Name  
14th Author's Affiliation ()
15th Author's Name  
15th Author's Affiliation ()
16th Author's Name  
16th Author's Affiliation ()
17th Author's Name  
17th Author's Affiliation ()
18th Author's Name  
18th Author's Affiliation ()
19th Author's Name  
19th Author's Affiliation ()
20th Author's Name  
20th Author's Affiliation ()
21st Author's Name  
21st Author's Affiliation ()
22nd Author's Name  
22nd Author's Affiliation ()
23rd Author's Name  
23rd Author's Affiliation ()
24th Author's Name  
24th Author's Affiliation ()
25th Author's Name  
25th Author's Affiliation ()
26th Author's Name / /
26th Author's Affiliation ()
()
27th Author's Name / /
27th Author's Affiliation ()
()
28th Author's Name / /
28th Author's Affiliation ()
()
29th Author's Name / /
29th Author's Affiliation ()
()
30th Author's Name / /
30th Author's Affiliation ()
()
31st Author's Name / /
31st Author's Affiliation ()
()
32nd Author's Name / /
32nd Author's Affiliation ()
()
33rd Author's Name / /
33rd Author's Affiliation ()
()
34th Author's Name / /
34th Author's Affiliation ()
()
35th Author's Name / /
35th Author's Affiliation ()
()
36th Author's Name / /
36th Author's Affiliation ()
()
Speaker Author-1 
Date Time 2026-06-18 09:30:00 
Presentation Time 25 minutes 
Registration for ICSS 
Paper # IA2026-14, ICSS2026-14 
Volume (vol) vol.126 
Number (no) no.73(IA), no.74(ICSS) 
Page pp.85-92 
#Pages 8 
Date of Issue 2026-06-10 (IA, ICSS) 

[Return to Top Page]

[Return to IEICE Web Page]
The Institute of Electronics, Information and Communication Engineers (IEICE), Japan