Detecting electromagnetic (EM) signals generated by electrical and electronic processes and equipment in the presence of large amounts of clutter and unknown obstacles or infrastructure along the channel presents a unique set of challenges, with the most critical being: Lack of signal design, Unknown channel conditions, and Potential active attempts to hide or obfuscate the signals. Recently, there have been significant progress in two areas that can significantly help with the main goal of detecting and classifying patterns of signal activity and discriminate signals of interest from background noise: Side channel analysis for integrity assessment and Deep Artificial Neural Networks. In terms of integrity assessment using side-channels, we have developed a technique called Power Fingerprinting (PFP) for monitoring execution integrity in digital devices. PFP monitors involuntary analog emissions, also known as side channels, coming directly from the processors to identify and classify specific patterns of known emissions from authorized execution and also to determine, via anomaly detection, when unauthorized operation is taking place. At the core, PFP technology identifies and classifies specific signal patterns and discriminates signals of interest from background noise, including co-located signals adding to the noise. Therefore, we can leverage the tools and techniques developed for PFP and scale them to help accurately detecting the operation of specialized electric or electronic equipment to advance our capabilities to monitor nonproliferation. PFP relies heavily on machine learning to learn the electromagnetic or direct power consumption patterns that result from the execution of legitimate or authorized software. One of the main tools we have leveraged for PFP machine learning are deep neural networks. “Deep learning” (DL) algorithms are uniquely suited to scenarios where the input data is not well-understood because they tend to require relatively little prior knowledge about the task at hand, and instead aim to learn all the necessary knowledge directly from data. Therefore, DL is well positioned to learn specific patterns of interest in unintended emissions from specific devices or processes. For this project, we propose to leverage PFP technology and deep learning frameworks to demonstrate the feasibility of discriminating specific signals of interest from background noise and detecting anomalies from known patterns. We plan on evaluating the performance of the proposed framework by testing direct connection and electromagnetic radiation in a testbed emulated with household appliances.
