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 2026-07-22 17:55
Evaluation experiment on the defensive effect of Differential Privacy against Model Inversion Attacks on transfer learning models
Aoki Koshiro, Kozo Okano, Shinpei Ogata (Shinshu Univ.), Shin Nakajima (NII) SS2026-15 KBSE2026-15
Abstract (in Japanese) (See Japanese page) 
(in English) Transfer learning is a machine learning technique that allows for the construction of highly accurate machine learning models with small amounts of data by repurposing the internal knowledge of a pre-trained learning model for a target task. Machine learning models have long been susceptible to the risk of sensitive data used for training being leaked due to privacy attacks. To mitigate this risk, DP-SGD, a learning algorithm based on differential privacy, has been used. On the other hand, given that transfer learning is currently the mainstream approach, it remains unclear whether DP-SGD is an effective
privacy protection method for transfer learning models. This report investigates the effectiveness of DP-SGD against model inversion attacks by creating multiple transfer learning models using SGD and DP-SGD and then applying two types of model
inversion attacks, GMI and KEDMI, to each model. The results showed that applying DP-SGD during transfer learning reduced attack accuracy, but also significantly reduced the model’s discrimination accuracy, resulting in a loss of model functionality. Even in cases where DP-SGD was applied during pre-training, additional experiments did not confirm any effect of DP-SGD. Therefore, it is considered that DP-SGD is not an effective model creation method for protecting training data from model inversion attacks.
Keyword (in Japanese) (See Japanese page) 
(in English) Machine Learning / Transfer Learning / SGD / DP-SGD / Model Inversion Attack / / /  
Reference Info. IEICE Tech. Rep., vol. 126, no. 125, SS2026-15, pp. 85-90, July 2026.
Paper # SS2026-15 
Date of Issue 2026-07-15 (SS, KBSE) 
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 SS2026-15 KBSE2026-15

Conference Information
Committee SS KBSE IPSJ-SE  
Conference Date 2026-07-22 - 2026-07-24 
Place (in Japanese) (See Japanese page) 
Place (in English)  
Topics (in Japanese) (See Japanese page) 
Topics (in English)  
Paper Information
Registration To SS 
Conference Code 2026-07-SS-KBSE-SE 
Language Japanese 
Title (in Japanese) (See Japanese page) 
Sub Title (in Japanese) (See Japanese page) 
Title (in English) Evaluation experiment on the defensive effect of Differential Privacy against Model Inversion Attacks on transfer learning models 
Sub Title (in English)  
Keyword(1) Machine Learning  
Keyword(2) Transfer Learning  
Keyword(3) SGD  
Keyword(4) DP-SGD  
Keyword(5) Model Inversion Attack  
Keyword(6)  
Keyword(7)  
Keyword(8)  
1st Author's Name Aoki Koshiro  
1st Author's Affiliation Shinshu University (Shinshu Univ.)
2nd Author's Name Kozo Okano  
2nd Author's Affiliation Shinshu University (Shinshu Univ.)
3rd Author's Name Shinpei Ogata  
3rd Author's Affiliation Shinshu University (Shinshu Univ.)
4th Author's Name Shin Nakajima  
4th Author's Affiliation National Institute of Informatics (NII)
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 2026-07-22 17:55:00 
Presentation Time 20 minutes 
Registration for SS 
Paper # SS2026-15, KBSE2026-15 
Volume (vol) vol.126 
Number (no) no.125(SS), no.126(KBSE) 
Page pp.85-90 
#Pages 6 
Date of Issue 2026-07-15 (SS, KBSE) 

[Return to Top Page]

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