@article{oai:ipsj.ixsq.nii.ac.jp:00239367, author = {加藤, 駿典 and 松井, 秀俊 and 寺田, 雅之 and Shunsuke, Kato and Hidetoshi, Matsui and Masayuki, Terada}, issue = {9}, journal = {情報処理学会論文誌}, month = {Sep}, note = {差分プライバシーは,数学的に証明可能なプライバシー保証を提供する指標であり,プライバシー保護研究の分野で大きな注目を集めている.差分プライバシーを提供するメカニズムの1つである圧縮メカニズムは,0を多く含むスパースなデータを観測行列を用いて1度低次元に圧縮し,付与すべきノイズを削減することで出力誤差を低減するが,感度の観点からノイズ削減の効果が低いと考えられる.本研究では,L1-制限等長性を満たすスパースな観測行列を用いることで,差分プライバシーを保証し,かつ出力誤差を低減する方法を提案する.提案手法の出力誤差を解析的に評価し,既存のメカニズムの出力誤差と比較する.また,メッシュ人口データを用いた評価実験を通じて,提案手法の性質や出力の傾向について考察し,提案手法の有効性を検証する., Differential privacy is a measure that provides mathematically provable privacy guarantees and has attracted considerable attentions in the field of privacy protection research. The compressive mechanism, one of the mechanisms that provide differential privacy, reduces the output error by compressing sparse data containing many zeros into low-dimensional data using a sensing matrix and reducing the noise to be added. However, there seems to be less effectiveness of the noise reduction from the viewpoint of the sensitivity. In this study, we propose a method for reducing the output error of the differential private mechanism by using a sparse sensing matrix that satisfies the L1 Restricted Isometry Property. We theoretically evaluate the output error of the proposed method and compare it with that of existing mechanisms. We also discuss the properties of the proposed method through experiments using mesh population data.}, pages = {1324--1338}, title = {圧縮メカニズム再考:スパースな観測行列を用いた圧縮センシングによる差分プライバシーの実現}, volume = {65}, year = {2024} }