| Presentation | 2023-02-22 Probabilistic Approach towards Theoretical Understanding for Adversarial Training Soichiro Kumano, Hiroshi Kera, Toshihiko Yamasaki, |
|---|---|
| PDF Download Page |  PDF download Page Link |
| Abstract(in Japanese) | (See Japanese page) |
| Abstract(in English) | In this paper, we provide the first theoretical analysis of the training dynamics of adversarial training of deep neural networks in a mean field view. To this end, we propose a new mean field theory, which allows us to study adversarial training from various aspects. Particularly, we derive tight upper bounds for adversarial loss with various norms, prove that networks are not adversarially trainable without shortcuts except for restrictive cases, show that adversarial training degrades the expressibility of a network, and the network width is an important factor to alleviate this. |
| Keyword(in Japanese) | (See Japanese page) |
| Keyword(in English) | Adversarial Training / Adversarial Examples / Mean Field Theory / Probabilistic Theory |
| Paper # | ITS2022-59,IE2022-76 |
| Date of Issue | 2023-02-14 (ITS, IE) |
| Conference Information | |
| Committee | IE / ITS / ITE-MMS / ITE-ME / ITE-AIT |
|---|---|
| Conference Date | 2023/2/21(2days) |
| Place (in Japanese) | (See Japanese page) |
| Place (in English) | Hokkaido Univ. |
| Topics (in Japanese) | (See Japanese page) |
| Topics (in English) | Image Processing, etc. |
| Chair | Kazuya Kodama(NII) / Masahiro Fujii(Utsunomiya Univ.) / Kenji Machida(NHK) / Hiroyuki Arai(Nippon Inst. of Tech.) / Hisaki Nate(Tokyo Polytechnic Univ.) |
| Vice Chair | Hiroyuki Bandoh(NTT) / Toshihiko Yamazaki(Univ. of Tokyo) / Kohei Ohno(Meiji Univ.) / Naohisa Hashimoto(AIST) / / Shogo Muramatsu(Niigata Univ) |
| Secretary | Hiroyuki Bandoh(KDDI Research) / Toshihiko Yamazaki(Nagoya Inst. of Tech.) / Kohei Ohno(Toyama Prefectural Univ.) / Naohisa Hashimoto(NIT, Tsuruoka College) / (Fukuoka Univ.) / Shogo Muramatsu(NHK) / (Hokkaido Univ.) |
| Assistant | Shunsuke Iwamura(NHK) / Shinobu Kudo(NTT) / Taishi Swabe(NAIST) / Keiji Jimi(Gunma Univ.) |
| Paper Information | |
| Registration To | Technical Committee on Image Engineering / Technical Committee on Intelligent Transport Systems Technology / Technical Group on Multi-media Storage / Technical Group on Media Engineering / Technical Group on Artistic Image Technology |
|---|---|
| Language | ENG-JTITLE |
| Title (in Japanese) | (See Japanese page) |
| Sub Title (in Japanese) | (See Japanese page) |
| Title (in English) | Probabilistic Approach towards Theoretical Understanding for Adversarial Training |
| Sub Title (in English) | |
| Keyword(1) | Adversarial Training |
| Keyword(2) | Adversarial Examples |
| Keyword(3) | Mean Field Theory |
| Keyword(4) | Probabilistic Theory |
| 1st Author's Name | Soichiro Kumano |
| 1st Author's Affiliation | The University of Tokyo(UTokyo) |
| 2nd Author's Name | Hiroshi Kera |
| 2nd Author's Affiliation | Chiba University(Chiba Univ.) |
| 3rd Author's Name | Toshihiko Yamasaki |
| 3rd Author's Affiliation | The University of Tokyo(UTokyo) |
| Date | 2023-02-22 |
| Paper # | ITS2022-59,IE2022-76 |
| Volume (vol) | vol.122 |
| Number (no) | ITS-384,IE-385 |
| Page | pp.pp.95-100(ITS), pp.95-100(IE), |
| #Pages | 6 |
| Date of Issue | 2023-02-14 (ITS, IE) |