Scaling Private Deep Learning with Low-rank and Sparse Gradients

• Item type: Trans(1)
• 公開日: 2023-10-19
• タイトル: Scaling Private Deep Learning with Low-rank and Sparse Gradients
• タイトル / 言語: en
• タイトル: Scaling Private Deep Learning with Low-rank and Sparse Gradients
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: [研究論文] deep learning, differential privacy, stochastic gradient decent
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_6501
• 資源タイプ: journal article
• 著者所属: Graduate School of Information Sciences and Technology, Osaka University
• 著者所属: LINE Corporation
• 著者所属: LINE Corporation
• 著者所属: Graduate School of Information Sciences and Technology, Osaka University
• 著者所属: Graduate School of Information Sciences and Technology, Osaka University
• 著者所属(英): en
• 著者所属(英): Graduate School of Information Sciences and Technology, Osaka University
• 著者所属(英): en
• 著者所属(英): LINE Corporation
• 著者所属(英): en
• 著者所属(英): LINE Corporation
• 著者所属(英): en
• 著者所属(英): Graduate School of Information Sciences and Technology, Osaka University
• 著者所属(英): en
• 著者所属(英): Graduate School of Information Sciences and Technology, Osaka University
• 著者名: Ryuichi, Ito; Seng, Pei Liew; Tsubasa, Takahashi; Yuya, Sasaki; Makoto, Onizuka
• 著者名(英): Ryuichi, Ito; Seng, Pei Liew; Tsubasa, Takahashi; Yuya, Sasaki; Makoto, Onizuka
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: Applying Differentially Private Stochastic Gradient Descent (DPSGD) to training modern, large-scale neural networks such as transformer-based models is a challenging task, as the magnitude of noise added to the gradients at each iteration scale with model dimension, hindering the learning capability significantly. We propose a unified framework, LSG, that fully exploits the low-rank and sparse structure of neural networks to reduce the dimension of gradient updates, and hence alleviate the negative impacts of DPSGD. The gradient updates are first approximated with a pair of low-rank matrices. Then, a novel strategy is utilized to sparsify the gradients, resulting in low-dimensional, less noisy updates that are yet capable of retaining the performance of neural networks. Empirical evaluation on natural language processing and computer vision tasks shows that our method outperforms other state-of-the-art baselines.; ------------------------------; This is a preprint of an article intended for publication Journal of; Information Processing(JIP). This preprint should not be cited. This; article should be cited as: Journal of Information Processing Vol.31(2023) (online); ------------------------------
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Applying Differentially Private Stochastic Gradient Descent (DPSGD) to training modern, large-scale neural networks such as transformer-based models is a challenging task, as the magnitude of noise added to the gradients at each iteration scale with model dimension, hindering the learning capability significantly. We propose a unified framework, LSG, that fully exploits the low-rank and sparse structure of neural networks to reduce the dimension of gradient updates, and hence alleviate the negative impacts of DPSGD. The gradient updates are first approximated with a pair of low-rank matrices. Then, a novel strategy is utilized to sparsify the gradients, resulting in low-dimensional, less noisy updates that are yet capable of retaining the performance of neural networks. Empirical evaluation on natural language processing and computer vision tasks shows that our method outperforms other state-of-the-art baselines.; ------------------------------; This is a preprint of an article intended for publication Journal of; Information Processing(JIP). This preprint should not be cited. This; article should be cited as: Journal of Information Processing Vol.31(2023) (online); ------------------------------
• 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: AA11464847
• 書誌情報: 情報処理学会論文誌データベース（TOD）; 巻 16, 号 4, 発行日 2023-10-19
• ISSN / 収録物識別子タイプ: ISSN
• ISSN / 収録物識別子: 1882-7799
• 出版者 / 言語: ja
• 出版者: 情報処理学会