Particulate Composites Aided Counterfeit Electronic Assembly Detection Using Generative Network Models

Vulnerabilities in the microelectronics supply chain warrant for a non-destructive counterfeit detection tool that enables authentication without expert intervention. Such tools must also be non-invertible (computationally hard to reverse engineer), and adaptable to different sub-applications and hardware setups. Convolutional Neural Networks (CNNs), which learn spatial characteristics, are commonly employed for such pattern recognition tasks. In this paper, we demonstrate the uniqueness of particle-filled polymeric composites for electronics assembly authentication. These composites, when embedded within or on a substrate, serve as physical taggants that reveal tampering or replacement. Physically valid synthetic microstructures mimicking particle-filled polymers are generated using a modified particle-packing algorithm and used to train neural networks. Our results show that networks, which train solely on classification, struggle with the open-set nature of the counterfeit data, often producing similar probability scores and are hard to classify. However, when reduced order representations of images are assumed gaussian and evaluated, the probability of authenticity can be precisely learned and used for robust classification, achieving accurate separation of true and counterfeit data. Networks which directly learn the probabilistic distribution of the true data and those that plot hyper sphere dimensionality reduction, similarly achieve near 100% accuracy on synthetic data, though their precision drops for counterfeit microstructures with closely related properties. These findings suggest that while classification based learning alone is not adequate, their trained latent representations provide reliable solutions for counterfeit detection in microstructural fingerprinting.