Human factors in secure and non-abusive machine learning systems

Mink, Jaron Maurice

Permalink

https://hdl.handle.net/2142/125593 Copy

Description

Title

Human factors in secure and non-abusive machine learning systems

Author(s)

Mink, Jaron Maurice

Issue Date

2024-07-10

Director of Research (if dissertation) or Advisor (if thesis)

Wang, Gang

Doctoral Committee Chair(s)

Wang, Gang

Committee Member(s)

• Redmiles, Elissa M
• Gunter, Carl
• Cobb, Camille

Department of Study

Siebel Computing &DataScience

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• machine learning
• deepfake
• security
• privacy
• adversarial machine learning
• trustworthy machine learning
• usable security
• human factors
• secure machine learning
• non-abusive machine learning
• human-ai interaction
• human-centric machine learning
• human-centric AI
• human-ml interaction

Abstract

Today, a significant portion of mission-critical work traditionally done by humans (e.g., driving cars, approving loans, medical triaging) is on the verge of being replaced by machine learning (ML). Historically, not considering human interactions with conventional software systems has led to significant harm. This is even more true for emerging ML systems as there is a lack of principled methods to construct safe and secure human-ML interaction paradigms. To prevent similar harm in ML-based systems, it is paramount that we understand vulnerabilities and apply safeguards now, while they are being designed and deployed. This dissertation investigates how the interaction of human factors and ML systems results in security implications via two perspectives. Specifically, this dissertation investigates how human factors can be exploited by ML-enabled abuse to reduce security in the context of deepfake deception (Chapter 3 and Chapter 4) and harnessed to improve the security of ML systems in the context of ML-enabled analysts tools (Chapter 5), and application of adversarial ML defenses (Chapter 6). In summary, these works show how human factors and perspectives contribute to the security of ML systems and that accounting for such interaction is necessary to protect from adversarial exploits.

Graduation Semester

2024-08

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/125593

Copyright and License Information

Copyright 2024 Jaron Mink

Owning Collections