東華大學圖書館 |

Language: English

Help

回圖書館首頁

手機版館藏查詢

Back

Switch To: Labeled | MARC Mode | ISBD

Linked to FindBook

Google Book

Amazon

博客來

On Statistical Robustness and Differential Privacy : = Theory and Algorithms.

Record Type:	Electronic resources : Monograph/item
Title/Author:	On Statistical Robustness and Differential Privacy :/
Reminder of title:	Theory and Algorithms.
Author:	Liu, Zheng.
Description:	1 online resource (162 pages)
Notes:	Source: Dissertations Abstracts International, Volume: 84-03, Section: B.
Contained By:	Dissertations Abstracts International84-03B.
Subject:	Statistics. -
Online resource:	http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=29326792click for full text (PQDT)
ISBN:	9798841759942

On Statistical Robustness and Differential Privacy : = Theory and Algorithms.
Liu, Zheng.

On Statistical Robustness and Differential Privacy :Theory and Algorithms. - 1 online resource (162 pages)

Source: Dissertations Abstracts International, Volume: 84-03, Section: B.

Thesis (Ph.D.)--The University of Wisconsin - Madison, 2022.

Includes bibliographical references

We study the statistical robustness and differential privacy, providing theories and algorithms. We first investigate the vulnerability of deep learning algorithms in the context of medical image segmentation and propose methods that may be adopted during training to make deep learning algorithms more robust. We show attack strategies based on Fast Gradient Sign Method (FGSM) can generate visually subtle perturbations to input images which severely affect segmentation result. We also show defense strategies, based on adversarial training and defensive distillation, show significantly improved robustness with respect to adversarial attacks. We then study robust estimation, where we propose Wasserstein contamination model and a W-GAN based robust estimator. We then present a general technique for upper-bounding the minimax rate of W-GAN-based estimators, and derive a general lower bound on the minimax rate via the modulus of continuity. The upper and lower bounds match in many estimation settings of interest, indicating that the W-GAN-based estimator achieves the optimal minimax rate.In the field of differential privacy, we study the differentially private estimation and synthetic dataset construction. For differentially private synthetic dataset construction, we propose a mechanism to efficiently release a synthetic dataset which could be used for various downstream applications. We give an Maximum Mean Discrepancy (MMD) bound between raw and released dataset. We also show this bound could be readily transformed into a bound for downstream analysis, e.g. principal component analysis (PCA), discriminant analysis and clustering, enabling the wide application of the mechanism. Then we consider differentially private estimation. We study the high dimensional linear regression model, using regularized M-estimator. We propose a differentially private regularized m-estimator via noisy composite gradient descent and show upper bound on the convergence rate of the private estimator.

Electronic reproduction.
Ann Arbor, Mich. :
ProQuest,
2023

Mode of access: World Wide Web

ISBN: 9798841759942Subjects--Topical Terms:

517247
Statistics.
Subjects--Index Terms:

Statistical robustnessIndex Terms--Genre/Form:

542853
Electronic books.

On Statistical Robustness and Differential Privacy : = Theory and Algorithms.
LDR:03266nmm a2200337K 4500 001 2362578
005 20231102121832.5
006 m o d
007 cr mn ---uuuuu
008 241011s2022 xx obm 000 0 eng d
020 $a 9798841759942
035 $a (MiAaPQ)AAI29326792
035 $a AAI29326792
040 $a MiAaPQ $b eng $c MiAaPQ $d NTU
100 1 $a Liu, Zheng. $3 2162356
245 1 0 $a On Statistical Robustness and Differential Privacy : $b Theory and Algorithms.
264 0 $c 2022
300 $a 1 online resource (162 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 84-03, Section: B.
500 $a Advisor: Zhang, Zhengjun.
502 $a Thesis (Ph.D.)--The University of Wisconsin - Madison, 2022.
504 $a Includes bibliographical references
520 $a We study the statistical robustness and differential privacy, providing theories and algorithms. We first investigate the vulnerability of deep learning algorithms in the context of medical image segmentation and propose methods that may be adopted during training to make deep learning algorithms more robust. We show attack strategies based on Fast Gradient Sign Method (FGSM) can generate visually subtle perturbations to input images which severely affect segmentation result. We also show defense strategies, based on adversarial training and defensive distillation, show significantly improved robustness with respect to adversarial attacks. We then study robust estimation, where we propose Wasserstein contamination model and a W-GAN based robust estimator. We then present a general technique for upper-bounding the minimax rate of W-GAN-based estimators, and derive a general lower bound on the minimax rate via the modulus of continuity. The upper and lower bounds match in many estimation settings of interest, indicating that the W-GAN-based estimator achieves the optimal minimax rate.In the field of differential privacy, we study the differentially private estimation and synthetic dataset construction. For differentially private synthetic dataset construction, we propose a mechanism to efficiently release a synthetic dataset which could be used for various downstream applications. We give an Maximum Mean Discrepancy (MMD) bound between raw and released dataset. We also show this bound could be readily transformed into a bound for downstream analysis, e.g. principal component analysis (PCA), discriminant analysis and clustering, enabling the wide application of the mechanism. Then we consider differentially private estimation. We study the high dimensional linear regression model, using regularized M-estimator. We propose a differentially private regularized m-estimator via noisy composite gradient descent and show upper bound on the convergence rate of the private estimator.
533 $a Electronic reproduction. $b Ann Arbor, Mich. : $c ProQuest, $d 2023
538 $a Mode of access: World Wide Web
650 4 $a Statistics. $3 517247
653 $a Statistical robustness
653 $a Differential privacy
655 7 $a Electronic books. $2 lcsh $3 542853
690 $a 0463
710 2 $a ProQuest Information and Learning Co. $3 783688
710 2 $a The University of Wisconsin - Madison. $b Statistics. $3 2101047
773 0 $t Dissertations Abstracts International $g 84-03B.
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=29326792 $z click for full text (PQDT)