Machine Learning and Privacy

Speaker: Prof. Pei-Yuan Wu
Date: 2018-12-06 (Thu) 10:30 - 12:00
Location: Auditorium 122 at CITI
Host: Chun-Shien Lu

Abstract:

  Private and sensitive data are commonly being collected and analyzed in machine learning applications. For instance, in biometric authentication, a user’s fingerprints, iris, or behavioral biometrics such as keystroke or mouse movements, are being collected for identity authentication, exempting the user’s burden of memorizing passwords or bringing smart cards.  However, the build-up of biometric authentication system requires collecting and analyzing bio-metrics from various users.  As a result, how to preserve privacy as well as preventing abusive usage of sensitive personal data, while at the same time enjoy the convenience and knowledge brought by deep learning, becomes an important issue.

  This talk aims to provide a broad overview over various security aspects in machine learning pipeline, including how security can be enhanced by applying machine learning to active authentication scheme, as well as security issues against attacks that use machine learning.  Threat models such as model inversion attacks, membership inference attacks, adversial example generation, as well as remedies including differential privacy, compressive privacy, and cryptographic approaches, will be introduced.

 

Bio:

  Private and sensitive data are commonly being collected and analyzed in machine learning applications. For instance, in biometric authentication, a user’s fingerprints, iris, or behavioral biometrics such as keystroke or mouse movements, are being collected for identity authentication, exempting the user’s burden of memorizing passwords or bringing smart cards.  However, the build-up of biometric authentication system requires collecting and analyzing bio-metrics from various users.  As a result, how to preserve privacy as well as preventing abusive usage of sensitive personal data, while at the same time enjoy the convenience and knowledge brought by deep learning, becomes an important issue.

  This talk aims to provide a broad overview over various security aspects in machine learning pipeline, including how security can be enhanced by applying machine learning to active authentication scheme, as well as security issues against attacks that use machine learning.  Threat models such as model inversion attacks, membership inference attacks, adversial example generation, as well as remedies including differential privacy, compressive privacy, and cryptographic approaches, will be introduced.