Black-Box Adversarial Attack for Math Formula Recognition Model

• Item type: SIG Technical Reports(1)
• 公開日: 2024-02-22
• タイトル: Black-Box Adversarial Attack for Math Formula Recognition Model
• タイトル / 言語: en
• タイトル: Black-Box Adversarial Attack for Math Formula Recognition Model
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: SIP2
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_18gh
• 資源タイプ: technical report
• 著者所属: Faculty of Science and Engineering, Doshisha University
• 著者所属: Faculty of Science and Engineering, Doshisha University
• 著者所属: Department of Information Engineering, University of Brescia
• 著者所属: Faculty of Science and Engineering, Doshisha University
• 著者所属(英): en
• 著者所属(英): Faculty of Science and Engineering, Doshisha University
• 著者所属(英): en
• 著者所属(英): Faculty of Science and Engineering, Doshisha University
• 著者所属(英): en
• 著者所属(英): Department of Information Engineering, University of Brescia
• 著者所属(英): en
• 著者所属(英): Faculty of Science and Engineering, Doshisha University
• 著者名: 名村, 晴人; 吉田, 正朋; アダミ, ニコラ; 奥田, 正浩
• 著者名(英): Haruto, Namura; Masatomo, Yoshida; Nicola, Adami; Masahiro, Okuda
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: Remarkable advances in deep learning have greatly improved the accuracy of image analysis. The progress of deep learning has significantly advanced the field of image analysis. However, concomitant with this advancement, the emergence of adversarial attacks aiming to deceive deep learning models has posed a significant challenge. While extensive research has focused on adversarial attacks, investigating such attacks targeting text recognition models, particularly optical character recognition (OCR), has remained largely unexplored. This paper presents a novel attack against a LaTeX OCR model designed to translate mathematical expression images into LaTeX code. Our method strategically exploits the characteristics of mathematical formula images by targeting filled areas containing text. This approach enables a computationally efficient attack on an image containing the mathematical formula.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Remarkable advances in deep learning have greatly improved the accuracy of image analysis. The progress of deep learning has significantly advanced the field of image analysis. However, concomitant with this advancement, the emergence of adversarial attacks aiming to deceive deep learning models has posed a significant challenge. While extensive research has focused on adversarial attacks, investigating such attacks targeting text recognition models, particularly optical character recognition (OCR), has remained largely unexplored. This paper presents a novel attack against a LaTeX OCR model designed to translate mathematical expression images into LaTeX code. Our method strategically exploits the characteristics of mathematical formula images by targeting filled areas containing text. This approach enables a computationally efficient attack on an image containing the mathematical formula.
• 書誌レコードID / 収録物識別子タイプ: NCID
• 書誌レコードID / 収録物識別子: AN10442647
• 書誌情報: 研究報告音声言語情報処理（SLP）; 巻 2024-SLP-151, 号 65, p. 1-5, 発行日 2024-02-22
• ISSN / 収録物識別子タイプ: ISSN
• ISSN / 収録物識別子: 2188-8663
• Notice: SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc.
• 出版者 / 言語: ja
• 出版者: 情報処理学会