Machine Learning Applied to Counterfeit Detection

The machine learning for counterfeit detection research program continues the successful Phase 1 feasibility study of applying machine learning to detect FPGA counterfeits. In Phase 1, Graf Research demonstrated the feasibility of implementing a machine learning based counterfeit detection platform for a single FPGA device and representing data characteristic of repackaged counterfeit devices.  The sensor development and FPGA characterization approaches demonstrated in the first phase are generally applicable across FPGA architectures with minimal alterations.  The second phase of this program will focus on improving and assessing the efficacy of that machine learning-based solution in classifying counterfeit devices by expanding the feasibility study to a full prototype characterization and machine-learning model-building, training, and classification solution. In addition, the second phase will demonstrate portability of the solution by identifying the features and characterization information necessary for expansion to an MPSoC with artificial intelligence IP cores.    This effort will result in a prototype platform for applying machine learning to detection of counterfeit devices.  The envisioned process will be broadly applicable throughout the supply chain using no external equipment.  The program focuses on producing a unique dataset resulting from non-destructive evaluation of devices to obtain aging-related data and employing machine learning to classify the data as indicating a device is genuine or counterfeit. Due to the non-destructive approach and lack of reliance on specialized external equipment, once a model is developed for a particular device and counterfeit class, cost of actual deployment into the supply chain is minimal.  The solution and demonstration that is ready for Phase 3 technical transition will serve as a non-destructive means for evaluating device authenticity at any point in the microelectronics for weapons systems supply chain.