Canada Spends Logo

Explore our Budget 2026 analysis

Grant: Discovery Grants Program - Individual

University of Victoria

$31,000.00

Differential Privacy Trade-offs in Machine Learning and Social Choice

FY 2024-25

Summary
With governments, businesses, and other organisations collecting an unprecedented amount of our individual information, the demand for better data privacy is higher than ever. Among a myriad of privacy definitions, differential privacy (DP) has become a "golden standard" for compliance to our new privacy legislations, and is currently used by organisations such as Apple, Google, and the United States Census Bureau. Yet, unlike other essential security measures like encryption and authentication, data privacy remains unstandardised, years after its popularisation. What impedes the standardisation, and thus a wider adoption of DP? One major obstacle is choosing privacy parameters. In DP, an algorithm's privacy is expressed via two parameters. They quantify the information leakage about individual persons in the data, if the output of this algorithm were released to the public. Smaller privacy parameters imply less leakage, resulting in better privacy. The difficulty in choosing privacy parameters then, lies in the fact that to achieve better privacy, an algorithm must sacrifice the amount of information in its output, causing it to lose some of its usefulness (or "utility"). A data analyst thus must balance the tradeoff between privacy and utility of an algorithm. More importantly, "utility" is subjective-in machine learning (ML) in particular, there exists multiple metrics of an algorithm's performance, each of which may have different tradeoffs to privacy parameters. With this state of affairs, it seems hopeless to attempt at standardising privacy, or to even compare private algorithms in some objective way. My research program proposes an alternative path to privacy standardisation-one which does not rely on subjective utility metrics, but an inherent tradeoff that exists between the privacy parameters of a differentially-private algorithm itself. We will first formally define, investigate, and create software for comparing private algorithms via the efficiency of their tradeoffs. Then, we will demonstrate the power of this tradeoff by studying private ML algorithms under this lens, to show that we can meaningfully compare such algorithms without resorting to subjective utility metrics. Finally, we will explore the generality of our proposed tradeoff. We will study the privacy tradeoffs of deterministic (in particular, voting) algorithms, whose analyses are only possible under variants of DP such as distributional DP. By collaborating with social scientists and surveying diverse groups of voters, we will use this tradeoff to quantify and control the privacy loss from outputting election data. My research program will simplify the choice of privacy parameters and private algorithms. Armed with a suite of new theoretical and software tools, data analysts and policy-makers will be encouraged to use DP for privacy compliance, leading to fewer privacy breaches in the future.
Program
Discovery Grants Program - Individual
Location
British Columbia, Canada
Recipient Type
Academia
Type
Grant
Source
NSERC's Awards Data

As published

Application id
RGPIN-2023-03642
Name nom
Lu, Yun
Department département
Computer Science
Organization id
7
Institution établissement
University of Victoria
Province en
British Columbia
Province fr
Colombie-Britannique
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2024
Competition year année de concours
2023
Award amount
31000
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
2700
Research subject group en
Information technology
Research subject group fr
Technologie de l'information
Research subject en
Information technology
Research subject fr
Technologie de l'information
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
Differential Privacy Trade-offs in Machine Learning and Social Choice
Keywords
differential privacy; privacy-preserving machine learning; privacy of election data; privacy parameters
Application summary
With governments, businesses, and other organisations collecting an unprecedented amount of our individual information, the demand for better data privacy is higher than ever. Among a myriad of privacy definitions, differential privacy (DP) has become a "golden standard" for compliance to our new privacy legislations, and is currently used by organisations such as Apple, Google, and the United States Census Bureau. Yet, unlike other essential security measures like encryption and authentication, data privacy remains unstandardised, years after its popularisation. What impedes the standardisation, and thus a wider adoption of DP? One major obstacle is choosing privacy parameters. In DP, an algorithm's privacy is expressed via two parameters. They quantify the information leakage about individual persons in the data, if the output of this algorithm were released to the public. Smaller privacy parameters imply less leakage, resulting in better privacy. The difficulty in choosing privacy parameters then, lies in the fact that to achieve better privacy, an algorithm must sacrifice the amount of information in its output, causing it to lose some of its usefulness (or "utility"). A data analyst thus must balance the tradeoff between privacy and utility of an algorithm. More importantly, "utility" is subjective-in machine learning (ML) in particular, there exists multiple metrics of an algorithm's performance, each of which may have different tradeoffs to privacy parameters. With this state of affairs, it seems hopeless to attempt at standardising privacy, or to even compare private algorithms in some objective way. My research program proposes an alternative path to privacy standardisation-one which does not rely on subjective utility metrics, but an inherent tradeoff that exists between the privacy parameters of a differentially-private algorithm itself. We will first formally define, investigate, and create software for comparing private algorithms via the efficiency of their tradeoffs. Then, we will demonstrate the power of this tradeoff by studying private ML algorithms under this lens, to show that we can meaningfully compare such algorithms without resorting to subjective utility metrics. Finally, we will explore the generality of our proposed tradeoff. We will study the privacy tradeoffs of deterministic (in particular, voting) algorithms, whose analyses are only possible under variants of DP such as distributional DP. By collaborating with social scientists and surveying diverse groups of voters, we will use this tradeoff to quantify and control the privacy loss from outputting election data. My research program will simplify the choice of privacy parameters and private algorithms. Armed with a suite of new theoretical and software tools, data analysts and policy-makers will be encouraged to use DP for privacy compliance, leading to fewer privacy breaches in the future.