AI news ·
How AJ Ketarkus is Strengthening AI to Resist Adversarial Attacks
AJ Ketarkus researched how AI models misinterpret altered inputs like obscured stop signs. His work helps train AI to better recognize key features and avoid critical errors.
Combatting Attacks on AI: Insights from Undergraduate Research
Artificial intelligence models are becoming more prevalent, but they can still make critical errors. One pressing question is how to prevent AI from misinterpreting data in ways that could lead to real-world consequences. This challenge motivated AJ Ketarkus, a sophomore studying computer sciences and electrical engineering, to focus his summer research on identifying and addressing vulnerabilities in AI systems.
Background and Motivation
AJ Ketarkus, originally from Ghana and raised in Madison, chose the University of Wisconsin–Madison for its strong computer science program and family legacy of success. Initially unaware of research opportunities, AJ’s participation in the Letters & Science Summer of Excellence in Research (LASER) program opened his eyes to the value of hands-on investigation during his undergraduate studies.
Research Focus: Inverse Adversarial Training through Optimal Transport
The core of AJ’s project involves a concept called “inverse adversarial training.” The idea is to work backward from potential AI errors to identify what might cause a model to misinterpret inputs. A practical example is a self-driving car encountering a stop sign that's been partially obscured by graffiti. The altered sign could confuse the AI, causing the vehicle to miss the stop instruction, which poses safety risks.
By analyzing these adversarial manipulations, the research aims to strengthen AI models so they recognize key features despite alterations. For example, training the AI to associate the red hexagonal shape with the stop command, regardless of missing letters, can prevent dangerous mistakes.
Why This Research Matters
Understanding how and why AI models fail is critical to improving their reliability. If researchers can pinpoint the triggers behind incorrect AI decisions, they can develop training methods that build resilience against those issues. This approach can enhance safety in applications like autonomous driving and beyond.
Lessons from the Research Experience
AJ shares that research is challenging, especially when starting out. There’s a steep learning curve filled with unknowns. However, collaboration with professors and fellow students created a supportive environment for growth. This teamwork allowed AJ to contribute meaningfully to the project as his understanding deepened.
Broader Impact on Professional Skills
Engaging in research provides more than technical knowledge. It teaches responsibility, problem-solving, and project management—skills that are valuable in any career path. Managing assignments independently and contributing to a collective goal prepares students for real-world challenges.
About the Researcher
AJ Ketarkus is a sophomore at the University of Wisconsin–Madison majoring in computer sciences and electrical engineering. Through the LASER program, he dedicated his summer to exploring strategies for defending AI models against adversarial attacks. AJ is also a scholar at the Center for Academic Excellence.
For those interested in expanding their expertise in AI and related fields, exploring targeted training programs can be valuable. Resources such as Complete AI Training's latest courses offer practical learning paths tailored to professionals in science and research.