Early-Warning Margins for Preamble-Directed Jamming in LoRa: A Dechirp-Domain Sensitivity Framework with...

Physical-layer jamming directed at the LoRa preamble may become detectable before substantial degradation of the acquisition structure is evident at the receiver. The practically relevant question is therefore not only whether an attack can eventually be detected, but how much jammer-power headroom a monitor provides before the selected preamble-degradation onset is reached. This paper introduces a detection-theoretic evaluation framework that expresses this pre-degradation warning capability in decibels. Three quantities are defined: the detection sensitivity floor, i.e., the smallest relative jammer level yielding a detection probability of 0.9 at a false-alarm probability of 10-2; the preamble-based link-degradation onset, i.e., the level at which the preamble symbol error rate increases by five percentage points relative to its jamming-free baseline; and the resulting early-warning margin separating the two thresholds. The resulting early-warning margin is a power-domain measure of jammer-power headroom and should not be interpreted as elapsed warning time. The framework is evaluated over 13 isolated impairment configurations, corresponding to 221 jammer-level operating points, and complemented by a three-step severity-graded composite-channel ladder that jointly combines noise, multipath, co-channel interference, and synchronization impairments, for more than 5×105 packet-level evaluations. Alongside a convolutional autoencoder and a one-class support vector machine using the same three-channel in-phase, quadrature, and magnitude preamble representation, we evaluate a training-light dechirp-domain peak-to-mean ratio statistic and a simple score-level fusion rule; under independent threshold calibration, the fusion rule preserves the separability of its constituents but, under fading, interference, and synchronization impairments, does not recover the fixed-operating-point sensitivity of the dechirp statistic. Under strong co-channel interference, the proposed scalar statistic reaches the target detection probability at -8.5 dB, whereas the autoencoder does not reach the target detection probability anywhere within the evaluated sweep, corresponding to a sensitivity advantage of more than 14 dB and showing that the observed loss of autoencoder sensitivity is detector-dependent rather than caused by an absence of discriminative information in the received preamble. No detector dominates across all isolated impairments: the dechirp statistic is particularly effective under co-channel interference and synchronization offsets, whereas the autoencoder remains advantageous under severe multipath fading. Some learned-detector margins become negative under moderate additive noise and under severe multipath fading, and the composite-channel ladder shows a progressive degradation of detection sensitivity as joint impairment severity increases. These results provide a detector-agnostic, physically interpretable yardstick for evaluating the pre-degradation sensitivity of LoRa physical-layer jamming monitors.