Defeating Cognitive Sensors

Cognitive sensors are conventional sensing hardware augmented by machine learning (ML). They are threats in the hands of our adversaries because they enable rapid, sensitive and extremely accurate signal classification and agile, customized, and potent responses. Current public demonstrations boast real-time, broadband radio frequency (RF) waveform classification with >99% accuracy at -15 dB SNR using little more than a GPU and software-defined radio. Against NAVAIR radars, where standoff capability is critical and adversary's electronic warfare (EW) has an R2 SNR advantage, cognitive sensors could disrupt the balance of power on the battlefield.To defeat an RF cognitive sensor, we must deny our adversaries access to ML. In this innovative Phase I effort, nSI will (a) make the first determination of whether transferable (black-box) attacks against ML-enabled RF signal classifiers are operationally feasible and (b) demonstrate robust, tactically relevant waveform mods which yield targeted errors against an ML-enabled radar signal classifier. These mods will be achievable on existing tactical H/W, account for channel effects and downconversion, and be imperceptible after matched filtering (no radar performance degradation). At the conclusion of this effort, any adversary that relies on ML for RF sensing will be subject to manipulation at our discretion.