P21_22: Cloud-based collaborative platform for defect identification and simulation
Topic Areas: Reverse engineering, Counterfeit, overproduction, IP theft detection and deterrence
Principal investigator: Dr. Sina Shahbazmohamadi, University of Connecticut
Co-Principal investigator:Dr. Pouya Tavousi, University of Connecticut
PI Email
Abstract:
Counterfeit and defective microelectronic parts continue to inflict significant damages on government, industry and society. This calls for finding effective ways to identify counterfeits. The current approaches involve acquisition of 2D and 3D images of the alleged part using a spectrum of microscopy tools, followed by having them assessed by a group of subject matter experts. This approach, nevertheless, entails two important shortcomings. Aside from the intensive computations needed for visualization, processing and analysis of the large microscopy data that we addressed in a previous project, due to lack of an objective measure for most classes of counterfeit, many defects are overlooked and even in some cases, they are falsely identified. Our proposed solution for addressing this challenge has three aspects. First, our solution provides a crowdsoucing mechanism, where a large groups of experts can contribute to the defect detection and analysis task through a web-based visualization and analysis platform. This will help form a collective insight through compilation of a data library. Second, the platform also enables simulation of different virtual scenarios such as different defective parts and/or normal variations within a part. In fact, the user can virtually introduce different classes of defects to a part and simulate a realistic images of the defective/counterfeit or healthy/authentic part through the platform. This has two advantages. First, it addresses the lack of data for training new inspection experts. Second, it provides fuel for training of machine learning algorithms for automating the inspection process. There is no limit on the amount of the data that can be generated using this method and more importantly the data is inherently labeled and thus, there is no need for manual labeling.