When the Rubbish Meets the Road: A Lesson About Data

Event Start
Event End
Building 3, Level 5, Room 5220.



AI/ML systems "learn" to make decisions based on the data with which 
they are trained.  Such systems are often used to make critical 
decisions in which mistakes can have  serious consequences -- e.g., 
systems for approving credit, job and college applications, digital 
forensic procedures, and computer-user authentication. In these kinds of 
applications AI/ML decision algorithms are tasked with distinguishing 
between legitimate and fraudulent or wrong behavior.

We show that minor degradations in as little as 1-2 percent of the 
training data can change decision outcomes by nearly 20 percentage 
points, wrongly reversing distinctions between legitimacy and 
fraudulence. In one real-world application – user authentication – data 
corruption was induced by USB keyboards injecting artifacts into the 
data, effecting an infidelity to the true signal. We illustrate how this 
phenomenon was discovered and validated.

Brief Biography

Roy Maxion is a Research Professor in computer science and machine 
learning at Carnegie Mellon University, where he is also the director of 
the Dependable Systems Laboratory.  His research has covered development 
and evaluation of highly reliable systems, human-computer interfaces, 
and automated detection, diagnosis and remediation of faulty or 
unanticipated events (anomalies) in many domains -- international 
banking, telecommunications networks, digital libraries, vendor help 
systems, semiconductor fabrication, process control, computer security, 
keystroke biometrics, camera ID forensics and others.  He is broadly 
experienced in experimental design and evaluation.

Dr. Maxion is a founding member of the NIST-supported, multi-university 
Center for Statistics and Applications in Forensic Evidence, whose 
mission is to build a scientifically and statistically sound foundation 
for formal and experimental analysis of forensic evidence.  He recently 
served as a member of the National Academy of Sciences committee on 
Future Research Goals and Directions for Foundational Science in 
Cybersecurity. He won an IEEE 2019 Test of Time Award (with Kevin 
Killourhy) for the 2009 experimental paper, "Comparing Anomaly Detection 
Algorithms for Keystroke Dynamics."  He is on the editorial boards of 
the International Journal of Machine Learning and IEEE Security & 
Privacy.  Dr. Maxion is an IEEE Fellow.

Contact Person