Artificial Intelligence for Detecting the Defect

Printed circuit boards (PCB) are found in nearly every electronic device imaginable. 

These boards are primarily responsible for connecting different components in a device and allowing constant communication between them. In other words, without the PCB, the device won’t work. 

Despite their importance, PCBs are becoming more susceptible to security risks, like malicious altering. 

“In most cases, these defects or alterations are hard to spot with the naked eye,” explained Mai Abdelhakim, principal investigator and assistant professor of electrical and computer engineering at the University of Pittsburgh Swanson School of Engineering. “Artificial intelligence and machine learning can detect defects from PCB images, but some components – like those with similar features or which are densely packed – aren’t clear in these images.”

A team of Pitt engineers led by Abdelhakim received a $900,000 grant from the Naval Nuclear Laboratory (NNL) to develop new machine-learning  algorithms that can detect and classify these potential threats inside a PCB. Abdelhakim will combine her expertise in cybersecurity and artificial intelligence with those of Co-Principal Investigator Samuel Dickerson’s in hardware and circuit design and Co-Principal Investigator Heng Ban’s in experimental systems in visual and infrared imaging and artificial intelligence in predictive maintenance.

Dickerson is an associate professor of electrical and computer engineering, vice chair for education and director of computer engineering undergraduate program at Pitt, while Heng Ban is the Richard K. Mellon professor of mechanical engineering and materials science, director of the Stephen R. Tritch Nuclear Engineering Program, and associate dean for strategic initiatives at Pitt.   

The team will use advanced infrared imaging and laser imagining tools available at the Multitask Thermophysics Laboratory at Pitt to produce thermal images of PCBs. 

“Infrared imaging allows us to obtain thermal images at different depths of the board so we can identify welds that can’t be examined with visible light,” Abdelhakim said.

Using these thermal images, the team will apply a deep learning technique to capture spatial correlations. 

To provide the models with a “fighting” mechanism, adversarial training will be added for robust detection and identification of defects in a PCB. 

“Adversarial training will enable us to boost the reliability of the machine learning models,” Abdelhakim said. 

The three-year project, “Artificial Intelligence for Hardware Security,” began in October 2023.