Issue link: https://read.uberflip.com/i/1520579
WHITE PAPER Achieving Full-Spectrum Dominance mrcy.com 8 emitter detection and deinterleaving, intelligent decision- making based on behavioral and environmental analysis, and AI-based countermeasure generation. Challenges of AI/ML in EW Although AI and machine learning (ML) can bring a vast array of capabilities to EW, they also bring a unique set of challenges such as reliability and reproducibility. This is due to the lack of transparency in neural-network-based algorithms. For traditional "rules-based" algorithms, the relationship between inputs and outputs is interpretable by a human subject matter expert, who can then construct a way to test the performance across all corner cases to ensure that the response of the algorithm to all possible inputs is known. Neural networks behave more like a black box and may sometimes behave counterintuitively when unconventional inputs are applied. An example of this is the "stop-sign sticker," in which a car 's ML system was trained to recognize stop signs. When small stickers were added to the stop sign, it fooled the car into thinking it was a 45-mph speed limit sign. Humans might feel comfortable accepting that their behavior, capabilities, strengths, and weaknesses are determined by both genetics and our experiences. For machines, however, we expect to be able to define a specification for a part number, and all part numbers should meet that specification, or we send it back to get fixed. For a cognitive EW system, the uniqueness of its mission necessarily limits the amount of training data, which implies it must learn in the field. But every system's field experience will be different, and its performance under standardized testing will almost certainly be different. Potential vulnerabilities such as this are the subject of adversarial ML, which in the context of EW, involves applying techniques that exploit ML vulnerabilities in a cognitive EW system to disrupt or degrade the system. EW is a cat- and- mouse game, one that is currently limited by human learning and development timescales. Cognitive EW will continue this, but limited only by machine processing and AI. Keeping up with AI/ML global advancements With global advancements in AI/ML increasing exponentially, it is critical to ensure that our test and training assets keep up with cognitive EW to meet current and future threats. Cognitive systems require more agility, awareness, and sophistication from pod-based airborne training jammers and live, virtual, and constructive (LVC) test simulators to provide sufficiently challenging targets to train against. Cognitive EW systems also require testing and evaluation performance beyond, the boundaries of current test equipment and procedures, so a separate cognitive system is needed for testing. Figure 3 illustrates how Mercury is integrating the capability to enable cognitive EW from chip scale to system level. COGNITIVE EW ARTIFICIAL INTELLIGENCE (AI) ADVANCEMENT AND THE EVOLUTION OF DRFM The most common use of a DRFM is to protect a platform by disrupting adversary radar tracking systems. A conceptually simple way to do this would be to blind the radar by emitting RF noise across the operating frequency range of every possible radar receiver, otherwise known as jamming. However, this approach is extremely inefficient because power is wasted by emitting at frequencies other than the operating frequency of the adversary 's radar receiver. For a jammer, maximizing output power is not just an academic problem; it also presents safety risks. A greater emitter power equates to a greater range, which in turn increases the time needed to make an avoidance maneuver—when a fraction of a second could mean the difference in avoiding an incoming threat. One way to make the jammer more efficient is to emit only at the operating frequency of the threat radar. A better way is to emit at the radar 's operating frequency and only when it is expecting a return pulse. Better still, the DRFM creates a return pulse that looks almost exactly like the true return, but is just slightly higher power, and then modifies this cover pulse to create a false target so that the adversary radar tracks it instead of the host platform. This technique, known as a range gate stealer (RGS), or range gate pull-off (RGPO), is part of the arsenal of techniques a DRFM has at its disposal. However, for sophisticated EW techniques such as these to be effective, the DRFM must know what the threat radar pulse looks like to find, capture, digitally modify, and emit that pulse. Currently, DRFMs know what to look for using a lookup table, known as a mission data file (MDF), which is loaded onto the DRFM before the mission. The MDF includes frequency, timing, and other characteristics of known threat radars. These are generated by teams of experts through analysis of signals intelligence data collections in a process that can take several months from the time an emitter is first seen in the field until an updated MDF is available to the DRFM. The need for cognitive EW to provide a faster MDF process Historically, radar updates were largely hardware-driven and the timeline to field new radars was several years, so an MDF update cycle of a few months was acceptable. However, newer software- defined radars can change via much faster software upgrades and, ultimately, with AI-enabled cognitive radars, could be reconfigured on the fly. To address the agility of modern radars, the DRFM must be able to detect a new enemy transmission, analyze it, and create an appropriate response on the fly. This is the goal of cognitive EW: To perform the entire MDF update cycle, previously performed by teams of human experts over a period of months, within seconds using AI algorithms. This is an incredibly lofty goal and will require advances in signal processing hardware, novel techniques for specific

