As the electromagnetic spectrum becomes increasingly crowded, it is critical that US ground force and tactical Signals Intelligence (SIGINT) and EW sensors are able to quickly and automatically scan large swaths of the RF spectrum and make sense of usage by private, commercial, civil and military entities operating a given area. Machine understanding of patterns of both routine and anomalous signal behavior would provide an indispensable advantage to operators monitoring and analyzing spectrum usage, allowing them to focus their efforts on only the most meaningful and actionable detections. The MUDSUM is an ML/AI system using a modular architecture which integrates deep signal detection, deep signal characterization, and a suite of pattern of life modeling methods, with domain relevant subject matter expertise. MUDSUM will be able to model dynamics in single emitters, multiple emitters of different classes, and additionally use context aware methods which model current behavior of the spectrum given the recent past. This multitiered and multiaspect method provides insight into typical and atypical behavior from multiple angles giving operators a large amount of visibility into the state and projected state of spectrum usage. MUDSUMs Phase I development will culminate in a proof of concept demonstration, evaluating the suite of signal detection, characterization, and behavior modeling tools.
Benefit: The MUDSUM systems capability will help to identify opportunities for blue force spectrum usage, and to focus attention on anomalous spectrum usage including those related to adversary actions. MUDSUM will include proven ML based techniques for monitoring large swaths of RF spectrum to identify distinct signals and to place them into several high level categories of signals (e.g., civilian comms such as WiFi or LTE or others, LPI, jammer, radar, etc.). These techniques perform highly efficient, near real time, coarse grained spectrum analysis with a high robustness to noise. MUDSUM will utilize deep characterization of discovered signals using proven techniques that combine deep learning and digital signal processing for joint modulation classification and signal parameter estimation. By deeply characterizing signals, MUDSUM is able to learn more nuanced and detailed pattern of normal activity and thereby more effectively discover anomalies. MUDSUM will leverage a suite of deep learning algorithms for anomaly detection and time series modeling are used to identify different kinds of anomalies, all of which can be operationally important. This includes signal level anomalies (e.g., this signal looks different than it normally does), aggregate anomalies (e.g., there are more of a particular kind of signal with a particular characteristic than usual), and context aware anomalies (e.g., there is an unusual number of a particular kind of signal given recent history and typical dynamics of the environment). MUDSUM will leverage templates detailing different kinds of anomalies of interest and the signal types, features, and characteristics that are relevant to that anomaly are provided by subject matter experts, and used to focus MUDSUM on only those patterns of behavior that are operationally relevant. For example, one template may describe the signal types and feature necessary to identify an adversary move to WARM mode, while another may describe the signal types and features that are involved with identifying opportunities for injecting blue force comms. Finally MUDSUM will be developed with a modular and scalable design, which can be adapted for use across a wide range of platforms and operational areas.
Keywords: Behavior Modeling, Behavior Modeling, Radar Communications, Deep Learning, anomaly detection, Radio Frequency, signal characterization, pattern of life, signal detection
