Search

This article proposes a residual network (ResNet)-based feature concatenated neural network model to predict the type of scenario the channel is under and the attribute of the predicted scenario with power delay profile (PDP) as the inputs. The generalized model structure consists of three blocks for feature extraction, scenario prediction, and attribute prediction, respectively. The PDP data is collected from a motherboard desktop environment under five different physical arrangement scenarios. Within each scenario, data is collected several times while varying a different physical attributes for each scenario. Two steps of data augmentation are applied to expand the size and to improve the resolution (difference between the neighboring attributes) of the measured dataset for the robust training and thorough evaluation of the proposed model. The proposed model is evaluated and compared with a multilayer perceptron (MLP)-based model on an expanded measured and averaged interpolated dataset. It is shown that both models perform very well on the expanded measured dataset with nearly 100% prediction accuracy on both scenarios and attributes. The MLP-based model suffers performance degradation on the averaged interpolated dataset with up to a 9% drop of classification accuracy on attribute prediction tasks, while our ResNet-based feature concatenated model performs equally in both scenarios. Feature activation mapping (FAM) and grad-class activation mapping (Grad-CAM) approaches are applied to provide visual explanations highlighting characteristics of the input PDP used for model decisions. FAM shows that the MLP-based model focuses on the multipath generated peaks of the PDP where some interpolated neighboring data points cannot be distinguished. The Grad-CAM shows that the proposed ResNet-based feature concatenated model performs better because it has strong attention not only on the multipath peaks but also on the valleys between those peaks which hold distinguishing information.

This paper introduces a novel method of non-destructively detecting incredibly elusive dormant hardware Trojans by selectively capturing hyperspectral backscattering measurements across the physical area of an integrated circuit. We propose a novel approach that pre-filters and actively samples an automatically selected set of the hyperspectral measurement space to significantly reduce measurement time while improving the ability to robustly detect dormant hardware Trojans. We demonstrate that our selective hyperspectral scanning approach can detect dormant hardware Trojans taking up as little as 0.03% of the circuit, which is up to 14 times smaller than prior work.

We show that a broad range of motherboard components can be classified based on their electromagnetic (EM) emanations using a lightweight convolutional neural network (CNN) architecture. Our random bootstrap sampling approach to cross-validation allows us to select robust models that achieve strong accuracy when classifying processor, memory, power, and ethernet integrated circuits (ICs) on internet of things (IoT) devices in two different excitation states. Our method also performs equal to or better than a k-Nearest Neighbors (k-NN) classifier baseline in all tested cases. We develop an anomaly detection procedure to flag the types of IC models not seen by the CNN in training. An analysis of signals passing through our end-to-end trained model allows us to generate insights about the discriminative features in the emanations of different components. We analyze how the convolution layers’ outputs pass through linear transformation layers to understand which of the input features are most important. Our results demonstrate that a CNN architecture can be used to classify ICs accurately using a diverse set of EM emanations while providing insights about the relevant discriminative features between components.