ML-Based EW Classifiers Waste Countermeasures on False Positives

Modern electronic warfare (EW) systems use machine learning to detect, classify, and respond to enemy radar and communications signals in real time. The problem is that the electromagnetic spectrum in a combat zone is extraordinarily congested — friendly radios, civilian signals, atmospheric noise, and enemy emissions all overlap. Current ML-based EW classifiers generate false positive rates of 10-30% in contested electromagnetic environments, meaning the system incorrectly identifies friendly or neutral signals as threats. Each false positive triggers a countermeasure response — jamming, chaff deployment, or evasive maneuver — that wastes limited resources and disrupts friendly operations. A ship that deploys chaff against a false alarm has fewer countermeasures available for a real attack. An aircraft that breaks formation to evade a non-existent missile loses tactical position. Over the course of a multi-day engagement, accumulated false positives degrade combat effectiveness more than the enemy's actual electronic attacks. This persists because EW training data is fundamentally scarce and non-representative. You cannot collect realistic adversary radar signatures without being in an actual conflict or conducting extremely expensive red-team exercises. Training data from peacetime exercises does not capture the density and complexity of wartime electromagnetic environments. Additionally, adversaries deliberately design their emissions to be ambiguous — making their radar look like civilian signals or friendly systems — which is specifically intended to exploit ML classifiers' weaknesses. The physics of the problem means that improving detection sensitivity inevitably increases false positives, and no amount of algorithmic sophistication can fully resolve signals that are intentionally designed to be confusing.