Unintended electromagnetic emissions, called EM emanations, can be exploited to recover sensitive information, posing security risks. Metal shielding, used by defense organizations to prevent data leakage, is costly and impractical for widespread use. This issue is particularly significant for IoT devices due to their sheer volume and varied deployment environments. Therefore, there is a research need for an automated detection method to monitor facilities and address data leakage promptly. To resolve this challenge, in the preliminary version of this work 1, a CNN-based detection method was proposed using HDMI cable emanations that provided ~95% accuracy up to 22.5m but had limitations due to training data. In this extended version, we augment the initial study by collecting and characterizing emanation data from IoT devices, everyday electronics, and cables. We propose a harmonic-based emanation detection method by developing a computational harmonic detection algorithm. The proposed method addresses the limitations of the CNN-based method and provides ~100% accuracy not only for HDMI emanation (compared to ~95% in the earlier CNN method) but also for all other tested devices and cables. Finally, it has also been tested in different environments to prove its efficacy in practical scenarios.

Traditional deep learning models incur high computational overhead (approximately 1-30W) and are unsuitable for Ultra-Low-Power (ULP) wearable systems like smart glasses. In contrast, TinyML architectures are power-efficient but less accurate. We propose distributing a neural network (NN) between a ULP wearable node and a resource-rich hub to maintain high accuracy and low power consumption for applications like human-machine vision. Output features from the node are transmitted to the hub via low-power communication, where the remaining network runs on traditional GPUs. We introduce a Figure of Merit (FoM) to determine the optimal NN distribution point and a customized multiplier unit for ULP operation. Achieving ULP of 284μW and 9mW for different Autoencoder (AE) networks, our approach is 700× lower in power consumption compared to traditional GPU implementation.

The transition of data and clock signals between high and low states in electronic devices creates electromagnetic radiation according to Maxwell's equations. These unintentional emissions, called emanation, may have a significant correlation with the original information-carrying signal and form an information leakage source, bypassing secure cryptographic methods at both hardware and software levels. Information extraction exploiting compromising emanations poses a major threat to information security. Shielding the devices and cables along with setting a control perimeter for a sensitive facility are the most commonly used preventive measures. These countermeasures raise the research need for the longest detection range of exploitable emanation and the efficacy of commercial shielding. In this work, using data collected from 3 types of commercial HDMI cables (unshielded, single-shielded, and double-shielded) in an office environment, we have shown that the CNN-based detection method outperforms the traditional threshold-based detection method and improves the detection range from 4 m to 22.5 m for an iso-accuracy of ~ 95%. Also, for an iso-distance of 16 m, the CNN-based method provides ~ 100% accuracy, compared to ~ 88.5% using the threshold-based method. The significant performance boost is achieved by treating the FFT plots as images and training a residual neural network (ResNet) with the data so that it learns to identify the impulse-like emanation peaks even in the presence of other interfering signals. A comparison has been made among the emanation power from the 3 types of HDMI cables to judge the efficacy of multi-layer shielding. Finally, a distinction has been made between monitor contents, i.e., still image vs video, with an accuracy of 91.7% at a distance of 16 m. This distinction bridges the gap between emanation-based image and video reconstruction algorithms.

Traditional methods of repairing a broken cable focus on restoring electrical connectivity and mechanical integrity, ignoring the electromagnetic aspects of it. Most of these repairing methods create a small monopole antenna as a byproduct which affects its electromagnetic compatibility (EMC). Switching activity in the transmitted signal through the wire creates an unintentional emission, called emanation, according to Maxwell's equations. This emanation is usually weak and suppressed to conform to EMC requirements. However, the monopole antenna of the repaired cable helps transmit it better, increasing the SNR of the emanation and extending its detection range significantly. This creates a serious security issue as emanations contain a significant correlation with the source signal and can be exploited for information extraction. In this work, the electromagnetic aspects of the broken cable repairing process have been explored in detail. We have applied the most commonly used cable repairing methods (twisting, soldering, and butt connector) to 3 types of widely used cables (USB, power, and HDMI cable) which are broken intentionally for experimental purposes. Collected data shows that the emanation SNR increases significantly due to the repairing process with −47 dBm power at a 20 cm distance. Although emanation power varies from cable to cable, it remains detectable even at >4 m distances. This strong emanation can penetrate through obstacles and remain detectable up to ~1$\mathbf{m}$distance through a 14 cm thick concrete wall. Along with exploring the vulnerability, a possible remedy, external metal shielding, has been explored in detail. This work exposes a new dimension of information leakage.

Stochastic variation of process parameters within a die and technology-limitation-driven variation from die-to-die give rise to unique distribution patterns for manufacturing process parameters. These patterns work as a process signature that is transferred from the device level to the system level through electrical circuits and can be used to make a distinction among the processes. In this work, we propose an in-situ manufacturing process technology distinction method, radio frequency process specific functions (RF-PSFs), that uses process-specific inherent properties of an IC manifested in the transmitted radio frequency signal. Among many desirable testing criteria, RF-PSF addresses the question of fabrication with the intended process technology. This information plays an important role in modern zero-trust architecture and IC clone detection, a counterfeiting method where the IC is manufactured using a different process. An RF transmitter with RF-DAC power amplifier for QPSK modulation has been designed and simulated in 14, 22, and 65 nm processes for five process corners (TT, FF, FS, SF, and SS) in Cadence. The simulated data have been processed in MATLAB. A multilayer perceptron (MLP), trained with the constellation data, provides an average accuracy of${\sim }90\%$for process distinction. Realizing that: 1) a higher order modulation will have even more process information and 2) we can harness the convolutional neural network’s (CNNs) improved capability on pattern recognition, we can feed image-like constellation plots to a CNN to get better and consistent performance. Using the baseband constellations for 64-QAM modulated data as images, we have achieved${\sim }100\%$accuracy with commonly used, pretrained CNN models (ResNet18, ResNet50, and GoogleNet) through transfer learning. The separation among five process corners within a process, termed intraprocess variation, is also analyzed. The effect of baseband sampling rate and ADC resolution, two practical limitations in RF systems, have been explored. An extensive study has been performed on the effect of a key design parameter at the RF circuit level, i.e., W/L or aspect ratio, leading to design insights, proper CNN selection, and some control parameters. This work establishes RF-PSF as a zero-power, zero-area overhead, and in-situ process distinction method.

Unlike traditional radio frequency device authentication which utilizes security keys in conjunction with a digital subsystem for verification, human voice communication involves probabilistic identification of a person based on his/her voice signatures and improves the detection probability over time. Inspired by voice-based human identification, we implement a novel method of augmenting trust during device detection and authentication, involving dynamic irregular clustering which exploits the unique nonidealities in IoT devices as physical signatures originated from Radio Frequency (RF) circuitry. The proposed method increases the confidence level of the classification as more data come in from a particular device, and is also able to detect new devices that do not fall into any of the previous clusters. Using 30 Xbee modules as transmitters, we show that our proposed method can detect a transmitter with > 95% sensitivity 100% with optimum parameters) using only 0.2 milliseconds of test data which makes it suitable for a very low latency communication system. Also, the incremental learning feature of the proposed method renders a gradual increase in sensitivity as more data are available from the transmitter end. The proposed method can provide an additional security layer in conjunction with the existing methods without adding any additional burden, which is extremely important for resource-limited asymmetric IoT nodes.