Improving the Performance of Deep Learning Image Classification Applied with Homomorphic Encryption

• Item type: Symposium(1)
• 公開日: 2023-10-23
• タイトル: Improving the Performance of Deep Learning Image Classification Applied with Homomorphic Encryption
• タイトル / 言語: en
• タイトル: Improving the Performance of Deep Learning Image Classification Applied with Homomorphic Encryption
• 言語: eng
• 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: Waseda University
• 著者所属: Waseda University
• 著者所属: Waseda University／NICT／RIKEN AIP
• 著者所属(英): en
• 著者所属(英): Waseda University
• 著者所属(英): en
• 著者所属(英): Waseda University
• 著者所属(英): en
• 著者所属(英): Waseda University / NICT / RIKEN AIP
• 著者名: Tianying, Xie; Hayato, Yamana; Tatsuya, Mori
• 著者名(英): Tianying, Xie; Hayato, Yamana; Tatsuya, Mori
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Ensuring privacy in deep learning applications is a critical issue in today’s society, and privacypreserving deep learning (PPDL) has attracted attention as a potential solution. Among various methods, PPDLs using Homomorphic Encryption (HE) are considered promising. However, while recent research has developed and evaluated PPDL algorithms based on HE, the achieved accuracy and latency require further improvements for practical application. In this study, we investigate the performance improvement of HEbased PPDL image classification by combining two techniques: the application of channel-level homomorphic encryption (CHE) and batch normalization (BN) with coefficient integration. Although these are widely used methods, specific algorithms and formulas have not been exhaustively detailed. A major contribution of this research is to provide comprehensive and reproducible descriptions of these techniques. In this study, we evaluate the implementation of CHE and BN using the Cheon-Kim-Kim-Song HE scheme, convolutional neural networks for deep learning, and the MNIST dataset. Furthermore, we compare the obtained results with the latest five neural network architectures. Experimental results show high accuracy (up to 99.32%) and short processing time (as low as 7.76 seconds), which is superior to existing research. Our approach suggests potential for supporting the design of more robust and flexible PPDLs.
• 書誌情報: コンピュータセキュリティシンポジウム2023論文集; p. 63-70, 発行日 2023-10-23
• 出版者 / 言語: ja
• 出版者: 情報処理学会