This new AI system detects counterfeit chips with high accuracy
The semiconductor industry, a $500 billion global market, is grappling with the shortage of new processors and an increase in counterfeit chips. This has inadvertently led to a $75 billion counterfeit chip market, posing substantial risks to sectors reliant on semiconductors, including aviation, communications, quantum computing, AI, and personal finance. To address this, Purdue University researchers have proposed an AI-based optical anti-counterfeit detection method for semiconductor devices named "Residual, Attention-based Processing of Tampered Optical Responses" (RAPTOR).
A novel innovation to counterfeit chip detection
The Purdue University team's novel approach, RAPTOR, identifies tampering by analyzing gold nanoparticle patterns embedded on chips. This method is designed to be robust under adversarial tampering features like malicious package abrasions, adversarial tearing, and compromised thermal treatment. The introduction of RAPTOR marks a significant shift from previous methods, that relied heavily on physical security tags incorporated into the chip functionality or packaging.
RAPTOR uses a deep-learning approach
To develop RAPTOR, the Purdue University team created a 10,000-image dataset of randomly distributed gold nanoparticles. These images were augmented from the original ones taken from a dark-field microscope. The researchers then clustered nanoparticle pattern pixel regions into local particle patterns and extracted their centers of mass. Finally, they generated Distance matrix Physical Unclonable Functions (PUFs) by evaluating all pairwise distances between these nanoparticle patterns.
It outperforms previous methods in counterfeit detection
The Purdue University team tested RAPTOR's anti-counterfeit capabilities by simulating tampering behavior in nanoparticle PUFs. This included both natural changes as well as malicious adversarial tampering. In these tests, RAPTOR showed the highest accuracy, correctly detecting tampering in 97.6% of distance matrices under worst-case tampering scenarios. This performance outstripped previous methods — Hausdorff, Procrustes, Average Hausdorff Distance — by 40.6%, 37.3%, and 6.4%, respectively.