Canada Spends Logo

Explorez notre analyse du budget 2026

Grant: Discovery Grants Program - Individual

University of Alberta

$29,000.00

Privacy in Machine Learning: Scalability and Limits

EF 2024-25

Résumé
Users are revealing, through actions online and through participation in business services, more information about themselves than ever before, and their behaviour is being collected and profiled. Profiles are then used, with varying degrees of success, for recommendation of products like films and books, targeting of advertisements, political campaigns and other services based on Machine Learning (ML). The level of data collection and profiling this entails clearly leave users vulnerable to privacy violations. The aim of privacy-preserving ML algorithms, the focus of my proposal, is to limit this vulnerability. Differential Privacy (DP) has been widely accepted as the gold standard definition of privacy because it allows rigorous mathematical guarantees of privacy. DP allows plausible deniability of an individual's presence in a database, which is one intuitive manner of understanding privacy.  DP is typically achieved by introducing perturbations, in the form of carefully calibrated random noise. Due to the addition of random noise, DP methods exhibit a tension between privacy and the accuracy of the model, often called utility. That is, as the level of privacy increases, the accuracy suffers. There have been many works that have proposed differentially private algorithms for common ML models such as logistic regression and deep learning. My research program is centered on the long-term objective of designing privacy-preserving machine learning algorithms for practical applications and real life datasets, addressing many gaps in the current landscape. The first such gap is that most current implementations assume datasets with independent data points, whereas in many real scenarios data points are linked and structured in graphs, such as social networks.  Employing standard methods of achieving DP implies adding larger noise due to the dependencies, sharply impacting model accuracy. Addressing graph structured data for private ML models will be a mid-term objective in my research. Second, data sharing between various entities in a privacy-preserving way has received less attention while it is a practical scenario. Many companies share data about their clients with partners, and the release of such datasets can cause significant privacy breaches. Generation of private synthetic datasets that can be shared will be second mid-term objective in my research program. The third gap I will address is the compatibility of privacy with fairness, another metric of AI ethics.   A third mid to long term objective of my research program is to study the interplay between privacy and fairness. If sensitive features used in fairness-promoting algorithms are also the ones to keep private, how is the assessment of fairness impacted by privacy-preserving methods? We need to go beyond sensitive features, as there are often surrogate features that contain the same information. Can we assure both fairness and privacy simultaneously?
Programme
Discovery Grants Program - Individual
Bénéficiaire
Emplacement
Alberta, Canada
Type de bénéficiaire
Academia
Type
Grant
Source
NSERC's Awards Data

Tel que publié

Application id
RGPIN-2021-02917
Name nom
Hegde, Nidhi
Department département
Computing Science
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
2021
Award amount
29000
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
1507
Committee name en
Computer Science
Committee name fr
Sciences informatiques
Area of application code
802
Area of application group en
Information and communication services
Area of application group fr
Services d'information et de télécommunications
Area of application en
Information systems and technology
Area of application fr
Systèmes d'information et technologie de l'information
Research subject code
9999
Research subject group en
Not available
Research subject group fr
Non disponible
Research subject en
Not available
Research subject fr
Non disponible
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
Privacy in Machine Learning: Scalability and Limits
Keywords
privacy; machine learning; differential privacy; fairness
Application summary
Users are revealing, through actions online and through participation in business services, more information about themselves than ever before, and their behaviour is being collected and profiled. Profiles are then used, with varying degrees of success, for recommendation of products like films and books, targeting of advertisements, political campaigns and other services based on Machine Learning (ML). The level of data collection and profiling this entails clearly leave users vulnerable to privacy violations. The aim of privacy-preserving ML algorithms, the focus of my proposal, is to limit this vulnerability. Differential Privacy (DP) has been widely accepted as the gold standard definition of privacy because it allows rigorous mathematical guarantees of privacy. DP allows plausible deniability of an individual's presence in a database, which is one intuitive manner of understanding privacy.  DP is typically achieved by introducing perturbations, in the form of carefully calibrated random noise. Due to the addition of random noise, DP methods exhibit a tension between privacy and the accuracy of the model, often called utility. That is, as the level of privacy increases, the accuracy suffers. There have been many works that have proposed differentially private algorithms for common ML models such as logistic regression and deep learning. My research program is centered on the long-term objective of designing privacy-preserving machine learning algorithms for practical applications and real life datasets, addressing many gaps in the current landscape. The first such gap is that most current implementations assume datasets with independent data points, whereas in many real scenarios data points are linked and structured in graphs, such as social networks.  Employing standard methods of achieving DP implies adding larger noise due to the dependencies, sharply impacting model accuracy. Addressing graph structured data for private ML models will be a mid-term objective in my research. Second, data sharing between various entities in a privacy-preserving way has received less attention while it is a practical scenario. Many companies share data about their clients with partners, and the release of such datasets can cause significant privacy breaches. Generation of private synthetic datasets that can be shared will be second mid-term objective in my research program. The third gap I will address is the compatibility of privacy with fairness, another metric of AI ethics.   A third mid to long term objective of my research program is to study the interplay between privacy and fairness. If sensitive features used in fairness-promoting algorithms are also the ones to keep private, how is the assessment of fairness impacted by privacy-preserving methods? We need to go beyond sensitive features, as there are often surrogate features that contain the same information. Can we assure both fairness and privacy simultaneously?