Presentation | 2020-03-03 Adversarial Attack against Neural Machine Translation Systems Takeshi Sakamoto, Tatsuya Mori, |
---|---|
PDF Download Page | PDF download Page Link |
Abstract(in Japanese) | (See Japanese page) |
Abstract(in English) | It has been widely known that systems empowered by neural network algorithms are vulnerable against an intrinsic attack named ``adversarial example'', which can be generated by adding small perturbations to the original inputs, aiming at fooling the systems. In this paper, we target Neural Machine Translation (NMT) and present attacks that change the meaning of sentences by adding small perturbations to the translated sentences (Adversarial text). This attack can intentionally control the nuance of meanings for documents such as contracts, products, reviews and postings to SNS by politicians or experts, which may play a vital role in making a decision. In this work, we adopt Google translate as a widely used NMT system and apply our attack using common sentences to study the effectiveness of the attack. We demonstrate that the meaning of sentences could be changed by 55% and the success rate is higher than the existing methods that target text classification applications. |
Keyword(in Japanese) | (See Japanese page) |
Keyword(in English) | Adversarial Example / Neural Machine Translation / Homoglyph |
Paper # | ICSS2019-89 |
Date of Issue | 2020-02-24 (ICSS) |
Conference Information | |
Committee | ICSS / IPSJ-SPT |
---|---|
Conference Date | 2020/3/2(2days) |
Place (in Japanese) | (See Japanese page) |
Place (in English) | Okinawa-Ken-Seinen-Kaikan |
Topics (in Japanese) | (See Japanese page) |
Topics (in English) | Security, Trust, etc. |
Chair | Hiroki Takakura(NII) |
Vice Chair | Katsunari Yoshioka(Yokohama National Univ.) / Kazunori Kamiya(NTT) |
Secretary | Katsunari Yoshioka(NICT) / Kazunori Kamiya(KDDI labs.) |
Assistant | Keisuke Kito(Mitsubishi Electric) / Toshihiro Yamauchi(Okayama Univ.) |
Paper Information | |
Registration To | Technical Committee on Information and Communication System Security / Special Interest Group on Security Psychology and Trust |
---|---|
Language | JPN |
Title (in Japanese) | (See Japanese page) |
Sub Title (in Japanese) | (See Japanese page) |
Title (in English) | Adversarial Attack against Neural Machine Translation Systems |
Sub Title (in English) | |
Keyword(1) | Adversarial Example |
Keyword(2) | Neural Machine Translation |
Keyword(3) | Homoglyph |
1st Author's Name | Takeshi Sakamoto |
1st Author's Affiliation | Waseda University(Waseda Univ.) |
2nd Author's Name | Tatsuya Mori |
2nd Author's Affiliation | Waseda University(Waseda Univ.) |
Date | 2020-03-03 |
Paper # | ICSS2019-89 |
Volume (vol) | vol.119 |
Number (no) | ICSS-437 |
Page | pp.pp.125-130(ICSS), |
#Pages | 6 |
Date of Issue | 2020-02-24 (ICSS) |