The global semiconductor market, valued at $500 billion, faces a growing problem: counterfeit chips. These fake chips pose serious risks, from potential malfunctions to unwanted surveillance, creating a $75 billion market that threatens industries like aviation, communications, and finance.
Traditional methods to combat counterfeit chips often rely on physical security tags or physical unclonable functions (PUFs), which leverage unique physical characteristics difficult to replicate. Optical PUFs, based on the distinct optical responses of random media, offer a promising solution due to their ease of fabrication and quick measurement. Nano-scale metallic optical systems, with their strong scattering response at optical wavelengths, have gained popularity. However, scaling up these systems and differentiating between intentional tampering and natural degradation remains challenging.
Researchers at Purdue University have developed a new approach called RAPTOR (Residual, Attention-based Processing of Tampered Optical Responses). This method uses deep learning to analyze gold nanoparticle patterns embedded on chips, identifying tampering even in challenging scenarios like malicious abrasions, thermal treatment, or tearing.
The research team first created a vast dataset of 10,000 images of randomly distributed gold nanoparticles, capturing their unique patterns. By analyzing the distances between these nanoparticles, they generated Distance Matrix PUFs, which serve as unique fingerprints for each chip. To test RAPTOR's ability to detect tampering, the researchers simulated various forms of intentional and natural alterations to the nanoparticle patterns.
RAPTOR's key innovation is its use of an attention mechanism. This allows the system to prioritize specific nanoparticle correlations across the original and tampered images, enabling a more accurate analysis of the changes. This information is then fed into a deep convolutional classifier that identifies tampering with remarkable accuracy.
The results were impressive: RAPTOR correctly detected tampering in 97.6% of cases under worst-case scenarios, outperforming existing methods by a significant margin. This demonstrates the power of deep learning in combating counterfeit chips.
This research signifies a significant step forward in using deep learning for PUF authentication. By achieving high accuracy even in complex real-world scenarios, RAPTOR paves the way for wider adoption of deep-learning-based anti-counterfeiting methods in the semiconductor industry.