Presentation 2024-01-17
Detecting Adversarial Examples using Filtering Operation Based on JPEG-Compression-Derived Distortion
Kenta Tsunomori, Minoru Kuribayashi, Nobuo Funabiki,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) Image classifiers based on convolutional neural networks are caused misclassification by adversarial perturbations. In this paper, we propose a method to use the difference images before and after applying the denoising filter to the input images for training the adversarial examples detection system. The proposed method employs the distortion signals modulated by the difference information of the images before and after JPEG compression as the denoising filter. Results of this research. The proposed method shows the adversarial examples detection accuracy of more than 98%.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) adversarial examples / denoising filter / Fine tuning / JPEG compression / Scaling
Paper # EMM2023-87
Date of Issue 2024-01-09 (EMM)

Conference Information
Committee EMM
Conference Date 2024/1/16(2days)
Place (in Japanese) (See Japanese page)
Place (in English) Tohoku Univ.
Topics (in Japanese) (See Japanese page)
Topics (in English) Sense of Presence, Universal Media, Digital Entertainment, etc.
Chair Michiharu Niimi(Kyushu Inst. of Tech.)
Vice Chair Kotaro Sonoda(Nagasaki Univ.) / Hyunho Kang(NIT, Tokyo)
Secretary Kotaro Sonoda(Hiroshima City Univ.) / Hyunho Kang(Osaka Inst. of Tech.)
Assistant Naofumi Aoki(Hokkaido Univ.) / Kazuaki Nakamura(Tokyo Univ. of Science)

Paper Information
Registration To Technical Committee on Enriched MultiMedia
Language JPN
Title (in Japanese) (See Japanese page)
Sub Title (in Japanese) (See Japanese page)
Title (in English) Detecting Adversarial Examples using Filtering Operation Based on JPEG-Compression-Derived Distortion
Sub Title (in English)
Keyword(1) adversarial examples
Keyword(2) denoising filter
Keyword(3) Fine tuning
Keyword(4) JPEG compression
Keyword(5) Scaling
1st Author's Name Kenta Tsunomori
1st Author's Affiliation Okayama University(Okayama Univ.)
2nd Author's Name Minoru Kuribayashi
2nd Author's Affiliation Tohoku University(Tohoku Univ.)
3rd Author's Name Nobuo Funabiki
3rd Author's Affiliation Okayama University(Okayama Univ.)
Date 2024-01-17
Paper # EMM2023-87
Volume (vol) vol.123
Number (no) EMM-332
Page pp.pp.38-43(EMM),
#Pages 6
Date of Issue 2024-01-09 (EMM)