PhD Comprehensive Examination Announcement of Ali Abbasi Tadi:"Privacy Preserving in Machine Learning "

 

Date: Monday July 18th 2022 

Time: 12:00 PM to 2:00 PM 

Meeting URL:    https://us06web.zoom.us/j/81199726498?from=addon 

Passcode: If interested in attending this event, contact the Graduate Secretary at csgradinfo@uwindsor.ca with sufficient notice before the event to obtain the passcode.

 

Machine learning (ML) is emerging as a vital asset in today’s world. Technology conglomerates are taking measures to provide ML services to clients with no ML background. This happened to the emergence of Machine Learning as a Service (MLaaS). While MLaaS reproduced a new era in ML, it inherently caused privacy breaches. High-tech companies are investigating solutions to preserve the privacy of their customers while keeping them off the complex data science knowledge. In this presentation, we explore various privacy threats in MLaaS as well as Federated Learning and provide the counter-measures in the literature for them. On top of this, we elaborate on membership inference attack as a privacy leakage. In addition, we investigate crypto graphic approaches (i.e. Homomorphic Encryption, secure processors, etc.), perturbation approaches (i.e. Differential Privacy and dimensionality reduction), and architectural approaches (i.e. Split Learning, knowledge transfer, etc.) as three main categories for mitigation of such attack. Moreover, open problems and challenges in this area will be presented. Overall, the state-of-the-art experiments show that preserving privacy would negatively affect either performance or accuracy of a ML model. Therefore, a trade-off between accuracy and privacy needs to be taken. 

 

Keywords: Machine Learning as a Service, Membership Inference attack, Differential Privacy, Cryptography  

 

External Reader: Dr. Ning Zhang 

Internal Reader: Dr. Luis Rueda 

Internal Reader: Dr. Saeed Samet  

Advisor(s): Dr. Dima Alhadidi