Transfer Learning Model For Terahertz VLSI Testing Technology

N. Akter, M.R. Siddiquee, J. Suarez, M. Shur, and N. Pala
Florida International University, Florida, United States

Poster stand number: T110

Keywords: Hardware security, THz response, integrated circuits, VLSI, failure detection, convolutional neural networks, data augmentation, transfer learning

The THz response images of the radio-frequency ICs (RFICs) allow distinguishing working and defective or faked circuits. These images are generated by measuring the response at the RFIC package pins by scanning the RFIC by the focused THz beam. We applied convolution neural network (CNN) and transfer learning models for THz testing of RFICs and increased the number of image datasets by applying the data augmentation processes. We trained the CNN model with and without the transfer learning approach and obtained the unsecure image dataset representing altered or damaged ICs. The unsecure ICs were damaged with a high voltage pulse. We also modified the original image data by augmenting the data with random noise (~20% to 200%). The trained model distinguished the secure and unsecure IC images with 98% accuracy.