Teaching a live jamming detector with hindsight

A detector reviewing a network recording can use what happened next to interpret a suspicious signal. A live detector cannot. A preprint originally published on 3 November 2026, at 05:14 UTC, explores whether reinforcement learning can transfer some of that hindsight into a detector that must make decisions as measurements arrive.

Learning from a detector that can look ahead

Radio jamming can disrupt legitimate communications, but its traces in network measurements vary with attack strength, burst length and network conditions. In mission-critical settings, missing an attack may be more costly than raising a false alarm. That makes recall—the share of actual jamming samples detected—an important objective, although false alarms still matter.

In the preprint CORA: Teacher-Guided Causal O-RAN Adaptation for Jamming Detection, Sneihil Gopal and Richard A. Rouil investigate a way to improve that trade-off without using ground-truth attack labels for training or deployment. The original publication timestamp, normalized to UTC, is 2026-11-03T05:14:00.000Z.

Their starting point is SEM-CS, a statistical detector that examines network measurements retrospectively. Its centered windows and retrospective processing use future context, making it unsuitable for live decisions. Replacing those operations with past-only calculations reduced its mean F2-score from 0.892 to 0.677 on the authors’ testbed corpus. F2 combines precision and recall while giving greater weight to recall.

CORA instead keeps the retrospective detector as an offline teacher. The goal is not to make future measurements available during operation, but to learn useful adjustments from historical recordings.

What the controller changes

The live detector processes nine network indicators, including signal power statistics, signal-to-noise ratio, received signal strength and the number of completed measurements in a report. It compares measurements with trailing statistics and combines several kinds of evidence into an alarm decision.

A reinforcement-learning controller adjusts three settings: sensitivity, the number of consecutive detections needed to raise an alarm, and the grace allowed to bridge brief gaps in supporting evidence. It can also end an active alarm under specified conditions.

Every five samples, the controller chooses an incremental action. A mask prevents settings from exceeding their bounds and limits how frequently configurations can change. These restrictions enforce stability rather than prescribing which direction the controller should move.

During training, rewards compare the live detector’s decisions with the teacher’s judgments on historical logs. Missing a teacher-positive sample receives a larger penalty than alarming on a teacher-negative sample. Additional penalties discourage fragmented alarms and excessive changes.

The authors test two learning algorithms: Proximal Policy Optimization, or PPO, and a Deep Q-Network, or DQN. At deployment, the trained policy is frozen and the teacher is removed. The controller then uses only current and past observations. Ground-truth labels are reserved for evaluation; the teacher’s judgments are a training signal, not verified attack annotations.

Higher recall, with a substantial alarm cost

The evaluation uses ten 5G O-RAN testbed logs totaling 34.7 hours. Each configuration is held out in turn while training uses the other nine. The held-out run requires no teacher, retraining or target-specific calibration. Each learner is trained with five random seeds.

Averaged equally across configurations, PPO increased recall from the fixed default detector’s 0.687 to 0.771—an 8.4-percentage-point gain—with improvement on every held-out configuration. That recall difference was statistically significant after correction for multiple comparisons.

The cost was substantial: false-alarm intervals increased from 1.98 to 24.93 per 1,000 samples, while precision fell from 0.963 to 0.735. Here, a false alarm is an alarm interval that overlaps no actual jamming event. PPO’s mean F2-score was 0.742 versus 0.711 for the default, but its F2 advantages over the principal static and learned comparators were not statistically significant.

DQN settled at a more conservative operating point. Its F2-score of 0.732 was comparable to the 0.734 of a static configuration selected using the teacher on training logs. DQN produced about 61% fewer false-alarm intervals than that static comparator, with lower recall.

Removing the teacher and training PPO with a reward based on causal evidence alone reduced recall on every configuration, supporting the importance of retrospective guidance in this experiment.

Where the evidence stops

CORA was not the best choice under every objective. With a strict budget of five false-alarm intervals per 1,000 samples, the default met the budget on nine configurations, PPO on none and DQN on one. In pooled sample-weighted results, the default’s F2-score of 0.836 exceeded both DQN’s 0.823 and PPO’s 0.801.

Transfer remained vulnerable to changing conditions. Training only on constant-gain attacks and testing on dynamic-gain attacks left both learners behind the default detector. The authors also acknowledge that all ten configurations were used during method development, so this was not an untouched validation set.

The paper maps offline training and live inference onto O-RAN’s two controller timescales; this architectural mapping should not be confused with demonstrated field deployment. Regime-aware teachers, explicit false-alarm budgets and processing individual devices separately remain future work. The reported contribution is a measured recall–false-alarm trade-off on a testbed, not a universally superior jamming detector.