Toward managing catastrophic AI risks

Mazeika, Mantas

Permalink

https://hdl.handle.net/2142/124375 Copy

Description

Title

Toward managing catastrophic AI risks

Author(s)

Mazeika, Mantas

Issue Date

2024-04-24

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

Forsyth, David

Doctoral Committee Chair(s)

Forsyth, David

Committee Member(s)

• Li, Bo
• Lazebnik, Svetlana
• Krueger, David

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• AI safety
• AI risk, robustness
• red teaming
• neural trojans
• trojan detection
• alignment
• model stealing

Abstract

Artificial intelligence (AI) has rapidly improved over the past decade, leading to widespread adoption of AI systems and demonstrating the potential for AI to greatly benefit society. However, as with any powerful new technology, AI introduces risks that must be managed to fully realize these benefits. Recent breakthroughs in the generality of AI systems have drawn increased attention to AI risks, including those of a potentially catastrophic nature. To help manage these anticipated risks, we take a defense in depth approach, combining different areas of AI safety research to address different aspects of AI risk. We present research on making AI systems more robust to adversarial influence, monitoring AIs for hidden behavior and trojans, enabling AIs to understand and adhere to human values, and finally addressing systemic problems to enable increased transparency.

Graduation Semester

2024-05

Type of Resource

Thesis

Copyright and License Information

Copyright 2024 Mantas Mazeika

Owning Collections