Training Convolutional Neural Networks (CNN) for Counterfeit IC Detection by the Use of Simulated X-Ray Images

Abstract

X-ray based non-destructive testing coupled with a deep learning approach in identifying counterfeit ICs is becoming state of the art. The features of the component under test such as copper lead frame and silicon chip are revealed by X-ray radiography, that forms the feature space to classify authentic or counterfeit component. The complexity in using deep learning algorithm is the dependence of convolutional neural networks performance on the large quantity of dataset for training. Obtaining these datasets consume an immense amount of time and work effort, which is prone to suffer from a change in the dimension of the base structure and material characteristics of the component under test. The aim is to construct a virtual X-ray simulation tool to generate synthetic radiographic images of the component under test and validate its effectiveness as training data for the convolutional neural network in identifying counterfeit component. The principle of perspective projection is used to compute the ray path length traversed through an object, and the interaction of a photon with the material is modeled using X-ray attenuation law. The tool is designed to import STL file with the possibility to assign material; for this purpose, the CAD model of the component under test and some counterfeits are developed through re-engineering steps. In total 2000 synthetic X-ray images of the CUT and 1500 X-ray images of counterfeit components are generated to train the CNN. An Image classification algorithm is constructed with VGG16 CNN model, trained with synthetic images and maximum prediction accuracy of 99.60% is attained.