Canada Spends Logo

Explore our Budget 2026 analysis

Grant: Discovery Grants Program - Individual

University of Waterloo

$41,000.00

FOUNDRY: Faster multi-party cOmputations UsiNg DiffeRential privacY

FY 2023-24

Summary
According to a 2019 survey by the Office of the Privacy Commissioner, 92% of Canadians question the protection of their privacy. At the same time, Canadian data assets were estimated to be worth over $200 billion according to PricewaterhouseCoopers. The Canadian society needs to reconcile important and socially beneficial machine learning applications with privacy concerns and fundamental techniques to overcome this dichotomy exist. Multi-party computations enable two or more parties to jointly compute a function, e.g., a machine learning classifier, while not revealing more about their data than the result of the function. Differential privacy, furthermore, ensures that changes in the input result in bounded changes in the output, such that inferences from the function result are limited. Both techniques are deployed in industry to protect privacy in machine learning applications. However, many applications of these techniques are still prevented by efficiency problems. For example, while machine learning inference using multi-party computations is efficient, training of machine learning models is not and out of reach for practical deployment. In this research, I propose to combine multi-party computations and differential privacy in novel ways to improve the performance of such privacy-preserving computations. My previous results and those of others indicate that this specific combination of multi-party computations and differential privacy is fruitful. We have been able to reduce the complexity of a multi-party computation for collecting key-value statistics from O(n^2) to O(n) and improve its practical running from days to seconds using differential privacy. In this application, I propose short-term objectives to extend this research and build differentially private multi-party computations for hash tables and maximum selection. As a long-term objective, I aim to combine these techniques to train private machine learning classifiers as such differentially private multi-party computations. Current purely cryptographic approaches fail at providing the necessary efficiency for such training. Differentially private classifiers, also known as privacy-preserving machine learning (PPML), are necessary to prevent inferences about the training data. Hence, we expect to obtain improved privacy protection and increased performance at the same time. As part of this research, I will train four graduate - 2 PhD and 2 Master's - and three undergraduate students in research and advanced development of security and privacy techniques for data science and machine learning. Many Canadian companies, including my industry sponsors, e.g., the Royal Bank of Canada (RBC), eagerly seek to employ students with this background. This research will also enable future privacy-preserving machine learning applications which address Canadians' concerns for their privacy while still use the data available to Canadian businesses such as RBC.
Program
Discovery Grants Program - Individual
Location
Ontario, Canada
Recipient Type
Academia
Type
Grant
Source
NSERC's Awards Data

As published

Application id
RGPIN-2023-03244
Name nom
Kerschbaum, Florian
Department département
David R. Cheriton School of Computer Science
Organization id
33
Institution établissement
University of Waterloo
Province en
Ontario
Province fr
Ontario
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2023
Competition year année de concours
2023
Award amount
41000
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
801
Area of application group en
Information and communication services
Area of application group fr
Services d'information et de télécommunications
Area of application en
Computer software
Area of application fr
Logiciel
Research subject code
2713
Research subject group en
Information technology
Research subject group fr
Algorithms
Research subject en
Technologie de l'information
Research subject fr
Algorithmes
Application title
FOUNDRY: Faster multi-party cOmputations UsiNg DiffeRential privacY
Keyword
Applied cryptography; Differential privacy; Privacy-preserving machine learning; Secure multi-party computations
Application summary
According to a 2019 survey by the Office of the Privacy Commissioner, 92% of Canadians question the protection of their privacy. At the same time, Canadian data assets were estimated to be worth over $200 billion according to PricewaterhouseCoopers. The Canadian society needs to reconcile important and socially beneficial machine learning applications with privacy concerns and fundamental techniques to overcome this dichotomy exist. Multi-party computations enable two or more parties to jointly compute a function, e.g., a machine learning classifier, while not revealing more about their data than the result of the function. Differential privacy, furthermore, ensures that changes in the input result in bounded changes in the output, such that inferences from the function result are limited. Both techniques are deployed in industry to protect privacy in machine learning applications. However, many applications of these techniques are still prevented by efficiency problems. For example, while machine learning inference using multi-party computations is efficient, training of machine learning models is not and out of reach for practical deployment. In this research, I propose to combine multi-party computations and differential privacy in novel ways to improve the performance of such privacy-preserving computations. My previous results and those of others indicate that this specific combination of multi-party computations and differential privacy is fruitful. We have been able to reduce the complexity of a multi-party computation for collecting key-value statistics from O(n^2) to O(n) and improve its practical running from days to seconds using differential privacy. In this application, I propose short-term objectives to extend this research and build differentially private multi-party computations for hash tables and maximum selection. As a long-term objective, I aim to combine these techniques to train private machine learning classifiers as such differentially private multi-party computations. Current purely cryptographic approaches fail at providing the necessary efficiency for such training. Differentially private classifiers, also known as privacy-preserving machine learning (PPML), are necessary to prevent inferences about the training data. Hence, we expect to obtain improved privacy protection and increased performance at the same time. As part of this research, I will train four graduate - 2 PhD and 2 Master's - and three undergraduate students in research and advanced development of security and privacy techniques for data science and machine learning. Many Canadian companies, including my industry sponsors, e.g., the Royal Bank of Canada (RBC), eagerly seek to employ students with this background. This research will also enable future privacy-preserving machine learning applications which address Canadians' concerns for their privacy while still use the data available to Canadian businesses such as RBC.