Grant: Discovery Grants Program - Individual
University of Alberta
$44,000.00
Novel Statistical Estimation and Inference under Local Differential Privacy
FY 2024-25
Summary
Personal data is widely used for various purposes, such as facial recognition, personalized advertising, medical trials, and recommendation systems. While sensitive personal information has potential benefits, it also has risks. Differential Privacy has provided a rigorous framework for controlled disclosure. Data distributions are measured by altering one data point in a query made on a dataset. The mathematically solid guarantee of DP and its variants is accompanied by concerns about a weak link, the trusted curator. Using Local Differential Privacy (LDP), users never lose control over sensitive information. Data is encoded and altered locally before being sent to an untrusted central server. In turn, this makes it harder to design valid local DP algorithms. Further statistical inference and development of corresponding theoretical properties are difficult.
We consider two statistics to address these challenges, namely quantile and cumulative distribution function (CDF) under Local Differential Privacy (LDP). Based on binary inquiries, an algorithm was developed to estimate population quantiles in the quantile setting. Using the newly developed tools, we will be able to solve the problem of detecting change points under LDP using relevant quantile differences. Furthermore, we will provide exponentially decreasing tail probability bounds under mild assumptions about the smoothness of the density functions in order to study non-asymptotic behavior under LDP. In the cumulative distribution function (CDF) setting, we propose a novel algorithm that exploits an unexpected connection between LDP and the current status problem, a classical survival data problem. It will be demonstrated that the proposed novel method achieves uniform and L2 error bounds when estimating the entire CDF curve. As the density of the grid increases, the error bound improves, exhibiting an asymptotic normal distribution. As the number of grids increases relative to the sample size, the error bound changes smoothly. Our constrained isotonic estimator is deterministically computed, eliminating the need for hyperparameters. The developed asymptotical properties will be used to generate bootstrap samples for inference. Also, we will explore the hypothesis testing functionality of CDF.
These novel methods provide timely, critical and extremely valuable statistical estimation and inference. Many data sets can be analyzed with the risk of handling sensitive information. As an example, diabetes research can provide long-term benefits to society. Grad students, undergraduates, and postdoctoral researchers will benefit from the use of statistical estimation and inference methods to gain valuable skills to prepare them for academic and industrial careers. For ease of dissemination, the algorithms will be implemented in R or Python and made available on GitHub.
Program
Discovery Grants Program - Individual
Recipient
Location
Alberta, Canada
Recipient Type
Academia
Type
Grant
Source
NSERC's Awards Data
As published
Application id
RGPIN-2024-04986
Name nom
Kong, Linglong
Department département
Mathematical and Statistical Sciences
Organization id
9
Institution établissement
University of Alberta
Province en
Alberta
Province fr
Alberta
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2024
Competition year année de concours
2024
Award amount
44000
Program id
RGPIN
Program name en
Discovery Grants Program - Individual
Program name fr
Programme de subventions à la découverte - individuelles
Group en
DISCOVERY RESEARCH
Group fr
RECHERCHE AXÉE SUR LA DÉCOUVERTE
Committee code
1508
Committee name en
Mathematics and Statistics
Committee name fr
Mathématiques et statistique
Area of application code
1207
Area of application group en
Advancement of knowledge
Area of application group fr
Avancement des connaissances
Area of application en
Information, computer and communication technologies
Area of application fr
Technologie de l'information, des ordinateurs et des réseaux de communications
Research subject code
3000
Research subject group en
Statistics and probability
Research subject group fr
Statistique et probabilité
Research subject en
Statistics and probability
Research subject fr
Statistique et probabilité
Field of research list names en
Not available
Field of research list names fr
Non disponible
Socioeconomic objective list names en
Not available
Socioeconomic objective list names fr
Non disponible
Application title
Novel Statistical Estimation and Inference under Local Differential Privacy
Keywords
statistical estimation and inference; local differential privacy; quantile estimate; cumulative distribution function; asymptotical properties; computational algorithms; self-normalization; non-asymptotic properties; relative difference; survival data analysis
Application summary
Personal data is widely used for various purposes, such as facial recognition, personalized advertising, medical trials, and recommendation systems. While sensitive personal information has potential benefits, it also has risks. Differential Privacy has provided a rigorous framework for controlled disclosure. Data distributions are measured by altering one data point in a query made on a dataset. The mathematically solid guarantee of DP and its variants is accompanied by concerns about a weak link, the trusted curator. Using Local Differential Privacy (LDP), users never lose control over sensitive information. Data is encoded and altered locally before being sent to an untrusted central server. In turn, this makes it harder to design valid local DP algorithms. Further statistical inference and development of corresponding theoretical properties are difficult.
We consider two statistics to address these challenges, namely quantile and cumulative distribution function (CDF) under Local Differential Privacy (LDP). Based on binary inquiries, an algorithm was developed to estimate population quantiles in the quantile setting. Using the newly developed tools, we will be able to solve the problem of detecting change points under LDP using relevant quantile differences. Furthermore, we will provide exponentially decreasing tail probability bounds under mild assumptions about the smoothness of the density functions in order to study non-asymptotic behavior under LDP. In the cumulative distribution function (CDF) setting, we propose a novel algorithm that exploits an unexpected connection between LDP and the current status problem, a classical survival data problem. It will be demonstrated that the proposed novel method achieves uniform and L2 error bounds when estimating the entire CDF curve. As the density of the grid increases, the error bound improves, exhibiting an asymptotic normal distribution. As the number of grids increases relative to the sample size, the error bound changes smoothly. Our constrained isotonic estimator is deterministically computed, eliminating the need for hyperparameters. The developed asymptotical properties will be used to generate bootstrap samples for inference. Also, we will explore the hypothesis testing functionality of CDF.
These novel methods provide timely, critical and extremely valuable statistical estimation and inference. Many data sets can be analyzed with the risk of handling sensitive information. As an example, diabetes research can provide long-term benefits to society. Grad students, undergraduates, and postdoctoral researchers will benefit from the use of statistical estimation and inference methods to gain valuable skills to prepare them for academic and industrial careers. For ease of dissemination, the algorithms will be implemented in R or Python and made available on GitHub.