Towards Practical Secure Automatic Speech Recognition

• Item type: Symposium(1)
• 公開日: 2018-10-15
• タイトル: Towards Practical Secure Automatic Speech Recognition
• タイトル / 言語: en
• タイトル: Towards Practical Secure Automatic Speech Recognition
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: Secure Two-party Computation，Automatic Speech Recognition，Deep Neural Network
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_5794
• 資源タイプ: conference paper
• 著者所属: University of Tsukuba
• 著者所属: University of Tsukuba
• 著者所属: University of Tsukuba／JST CREST／RIKEN AIP Center
• 著者所属(英): en
• 著者所属(英): University of Tsukuba
• 著者所属(英): en
• 著者所属(英): University of Tsukuba
• 著者所属(英): en
• 著者所属(英): University of Tsukuba / JST CREST / RIKEN AIP Center
• 著者名: Jun-jie, Zhou; Wen-jie, Lu; Jun, Sakuma
• 著者名(英): Jun-jie, Zhou; Wen-jie, Lu; Jun, Sakuma
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Smart speakers are become popular thanks to the improvements of automatic speech recognition systems. However, the always-on system raises privacy problems. Moreover, we imagine in the future that the service provider may change the smart speakers to be active that the speakers will offer user-interested/user-aimed information according to the user conversations without being firstly evoked, which brings the further privacy concerns. In this study, we address this privacy issue and discover the feasibility of realizing such active smart speakers. Specifically, we mainly investigated the computation cost of using the state-of-the-art multiplication triples generation protocols to prepare for three different speech recognition models and give detailed results. The results show that the DNN model is the most practical secure ASR model at present with existing protocols because it only needs 6.35 hours for preparing private evaluation in the remaining 17.65 hours of a day.
• 書誌レコードID / 識別子タイプ: NCID
• 書誌レコードID / 関連識別子: ISSN 1882-0840
• 書誌情報: コンピュータセキュリティシンポジウム2018論文集; 巻 2018, 号 2, p. 623-630
• 出版者 / 言語: ja
• 出版者: 情報処理学会