Wed, Sep 07, 2022 @ 10:00 AM - 11:30 AM
PhD Candidate: Tiantian Feng
Title: Towards designing trustworthy human-centered machine intelligence: A privacy-enhancing perspective
Shrikanth Narayanan, Professor Laurent Itti, Professor Aiichiro Nakano, Professor Kristina Lerman, and Professor Morteza Dehghani (external)
Human-centered machine intelligence has revolutionized many leading domains ranging from transportation and healthcare to education and defense, profoundly changing how people live, work, and interact with each other. Human-centered ML systems often require acquiring and processing multimodal data from people in sensitive environments and contexts, such as homes, workplaces, hospitals, and schools. These systems also utilize state-of-art machine learning (ML) algorithms to achieve a deeper understanding of human conditions, such as state, trait, and interactions, which provide possibilities to create technologies that increasingly support and enhance human experiences. Despite the promises human-centric ML systems can deliver, they have drawn substantial concerns in recent years on their potential adverse impact on society, forcing researchers to design trustworthy ML systems to address fairness, accountability, privacy, and related topics before an ML system can be deployed widely. In particular, these systems create critical risks in potentially leaking sensitive information that people might want to keep private. The sensitive information can be individual attributes (e.g., age, gender), states (e.g., health, emotions), or biometric fingerprints. In this thesis proposal, I focus on exploring privacy risks and privacy-enhancing computation in modern human-centered ML systems. I will start the thesis proposal by introducing recent efforts that utilizes human-centered machine intelligence to capture meaningful human behavioral patterns. Then I introduce a series of my recent works on privacy attacks, such as attribute inference attacks and label inference attacks, which motivate us to develop mitigation strategies. I will then present a proposal that aims to extend our privacy-enhancing computation to other privacy attacks, like membership inference attacks and data reconstruction attacks. I also propose to study the impact of privacy-enhancing computing on other aspects of trustworthy computing, like fairness. The proposal will demonstrate the importance of privacy-enhancing computation, especially in designing trustworthy human-centered machine intelligence.
WebCast Link: https://usc.zoom.us/j/98609209991
Audiences: Everyone Is Invited
Contact: Lizsl De Leon