网络异构性下的 O-RAN 原生干扰检测

# O-RAN-Native Jamming Detection Under Network Heterogeneity
Source: [https://arxiv.org/html/2610.20816](https://arxiv.org/html/2610.20816)
Fernando J\. Cintrón⋆Richard A\. Rouil⋆Affiliation:‡PREP Associate, National Institute of Standards and Technology \(NIST\), USAAffiliation:†Department of Physics, Georgetown University, USAAffiliation:⋆National Institute of Standards and Technology \(NIST\), USA\.Affiliation:Emails: \{sneihil\.gopal, fernando\.cintron, richard\.rouil\}@nist\.gov

###### Abstract

Radio\-frequency jamming threatens cellular network operations in contested environments\. The Open Radio Access Network \(O\-RAN\) architecture exposes per\-User Equipment \(UE\) telemetry through the E2 interface, enabling Key Performance Indicator \(KPI\)\-based jamming detection without proprietary hardware\. However, operational deployment demands detection methods that generalize across heterogeneous networks\. We propose a Statistical Ensemble Method \(SEM\) combining adaptive Median Absolute Deviation \(MAD\)\-based baseline detection, count\-based anomaly detection, and abrupt change detection through a weighted voting system\. SEM operates in two modes: a cold\-start mode requiring no labeled data for immediate deployment, and a supervised mode where parameters are refined offline via grid search as labeled data accumulates\. We evaluate SEM against five machine learning \(ML\) classifiers and five deep learning \(DL\) sequence models using approximately𝟐𝟓\\mathbf\{25\}hours of testbed data comprising logs from eight heterogeneous 5G environments, employing both single\-environment and multi\-environment leave\-one\-out \(LOO\) validation\. Our analysis indicates that under conditions not encountered during training \(distribution shift\), learning\-based models tend to lose recall and miss jamming events, whereas SEM, even in cold\-start mode, maintains detection capabilities at the cost of false alarms\. Furthermore, SEM’s parameter optimization can be mapped to O\-RAN’s Non\-RT RIC as an rApp for continuous refinement, allowing operators to transition to learning\-based alternatives once corpus coverage is sufficient\.

## IIntroduction

Radio\-frequency jamming poses a growing threat to military and civilian cellular network operations\. By overwhelming legitimate transmissions with broadband interference, a jammer can suppress reference signal measurements, degrade handover reliability, and cause complete service outages\. These disruptions can have serious consequences forcommand\-and\-control \(C2\)communications and critical infrastructure in contested environments\[[1](https://arxiv.org/html/2610.20816#bib.bib11)\]\. TheOpen Radio Access Network \(O\-RAN\)architecture offers an effective platform for jamming detection by enabling the monitoring of per\-User Equipment \(UE\)telemetry, includingReference Signal Received Power \(RSRP\)statistics,Signal\-to\-Noise Ratio \(SNR\), andReceived Signal Strength Indicator \(RSSI\)\. This data is made available through standardized E2 interface reports, which can be accessed by detection applications in theRAN Intelligent Controller \(RIC\)without the need for proprietary instrumentation\[[2](https://arxiv.org/html/2610.20816#bib.bib12),[3](https://arxiv.org/html/2610.20816#bib.bib1)\]\.

Machine learning \(ML\)classifiers anddeep learning \(DL\)sequence models have shown strong performance inKey Performance Indicator \(KPI\)\-based jamming detection during controlled evaluations\[[4](https://arxiv.org/html/2610.20816#bib.bib5),[5](https://arxiv.org/html/2610.20816#bib.bib2),[6](https://arxiv.org/html/2610.20816#bib.bib6)\]\. However, their effectiveness is constrained by two major challenges, particularly in military deployments\. The first is*distribution shift*: a model trained in one environment learns signal characteristics specific to that environment and may struggle when deployed under different operating conditions\. The second is the*cold\-start problem*: learning\-based models require labeled attack data before they can be trained for deployment\[[7](https://arxiv.org/html/2610.20816#bib.bib3),[8](https://arxiv.org/html/2610.20816#bib.bib9)\]\. In a new deployment, such labeled data is typically unavailable, meaning that detection capabilities cannot be established until sufficient examples are collected and annotated\.

To address these limitations, we propose aStatistical Ensemble Method \(SEM\)\. The core detectors ofSEMare based on signal characteristics and employ well\-defined default parameters, requiring no labeled data to begin operating\. When labeled data becomes available, secondary parameters can be refined offline through grid search\. This yields a two\-mode design: a*cold\-start mode*for immediate detection and a*supervised mode*for subsequent parameter refinement\. We evaluateSEMagainst fiveMLclassifiers and fiveDLsequence models using approximately2525hours of testbed data comprising eight heterogeneousfifth\-generation \(5G\)logs\. The evaluation considers both single\-environment and multi\-environmentleave\-one\-out \(LOO\)validation\. The logs vary inUEcount, recording duration, jammer level, and burst pattern\.

Across these heterogeneous conditions, the results reveal three main findings\. First, the advantage ofSEMis not fully reflected by its mean F1\-score, which balances precision and recall\. Instead, its strength lies in how it behaves under challenging conditions\. Under distribution shift, learning\-based models can lose recall and miss jamming events, whereasSEMmaintains detection at the cost of additional false alarms\. On the most challenging log, where jamming burst durations vary randomly, single\-environmentMLandDLachieve recalls of only0\.290\.29and0\.470\.47, respectively, whereasSEMmaintains a recall of0\.960\.96even in cold\-start mode\. Second, multi\-environment training substantially improves the performance of learning\-based models compared with single\-environment training\. Specifically, the F1\-score forMLincreases from0\.9110\.911to0\.9620\.962, while that ofDLincreases from0\.8970\.897to0\.9560\.956\. However, this improvement depends on the training corpus sufficiently representing the target deployment conditions, which cannot be guaranteed during initial rollout\. Third, the results indicate that corpus coverage, rather than hyperparameter configuration alone, is a primary constraint on the generalization of learning\-based methods across heterogeneous environments\.

The dual\-mode design ofSEMaligns naturally with theO\-RANRICarchitecture\. In cold\-start mode, SEM operates as aneXtended application \(xApp\)on the near\-real\-time \(RT\)RIC, receiving liveKPItelemetry through the E2 interface and allowing detection processing to begin immediately upon activation\. As labeled incidents accumulate, the non\-RTRICcan host a parameter\-optimizationRAN Application \(rApp\)that refinesSEM’s parameters through grid search\. The optimized parameters can then be delivered through the A1 interface, allowingSEMto transition to supervised mode without interrupting live detection\. Once the labeled dataset sufficiently represents the target deployment conditions, therAppcan assess whether a learning\-based alternative achieves performance comparable toSEMon validation data, allowing operators to determine whether a transition is appropriate\. This deployment approach addresses the cold\-start barrier that can hinder learning\-based methods in new deployments\.

The remainder of this paper is organized as follows\. Section[II](https://arxiv.org/html/2610.20816#S2)reviews related work\. Section[III](https://arxiv.org/html/2610.20816#S3)describesSEMand the learning\-based models\. Section[IV](https://arxiv.org/html/2610.20816#S4)presents the experimental setup\. Section[V](https://arxiv.org/html/2610.20816#S5)reports the results\. Section[VI](https://arxiv.org/html/2610.20816#S6)discussesO\-RANintegration, and Section[VII](https://arxiv.org/html/2610.20816#S7)concludes the paper\.

## IIRelated Work

Jamming detection approaches span signal\-level analysis, threshold\-based methods, andKPI\-based statistical detection\. Early work established that jamming produces characteristic degradation signatures \(RSRPdrops,SNRvariations, andRSRPcount loss\) observable without physical\-layer access\[[1](https://arxiv.org/html/2610.20816#bib.bib11)\]\. Threshold\-based methods require manual calibration per site and do not adapt to changing conditions\[[9](https://arxiv.org/html/2610.20816#bib.bib4)\], while statistical ensemble and time\-series approaches have shown promise for cellular anomaly detection\[[10](https://arxiv.org/html/2610.20816#bib.bib8),[11](https://arxiv.org/html/2610.20816#bib.bib14)\]\. More recently,KPI\-based detection has been demonstrated in5GNew Radio \(NR\)contexts\[[12](https://arxiv.org/html/2610.20816#bib.bib7),[13](https://arxiv.org/html/2610.20816#bib.bib10)\]and withinO\-RANarchitectures\[[2](https://arxiv.org/html/2610.20816#bib.bib12),[3](https://arxiv.org/html/2610.20816#bib.bib1),[14](https://arxiv.org/html/2610.20816#bib.bib13)\]\.

MLandDLapproaches have increasingly been applied to jamming and interference detection, including neural network\-based detection\[[8](https://arxiv.org/html/2610.20816#bib.bib9)\], hybrid classifiers for known and unknown attacks\[[5](https://arxiv.org/html/2610.20816#bib.bib2)\], Bayesian inference\-assisted classification in5GNR\[[4](https://arxiv.org/html/2610.20816#bib.bib5)\], and spectrogram\-based supervised and unsupervised detection\[[6](https://arxiv.org/html/2610.20816#bib.bib6)\]\. However, many evaluations are conducted under relatively homogeneous conditions, which can overestimate deployment performance by limiting the distribution shifts encountered during evaluation\. Supervised learning\-based approaches also require labeled attack data before deployment, creating a barrier in environments with no prior jamming history\.

Our work addresses both limitations\. We benchmark under both single\-environment and multi\-environmentLOOvalidation across eight heterogeneous5Glogs, directly characterizing distribution shift across deployment conditions\. By evaluating bothSEMand learning\-based methods under identical conditions, this paper provides, to the best of our knowledge, the first comparison that isolates the contribution of the detection architecture from that of labeled data\.

## IIIMethodology

This section describes the proposedSEM,MLclassifiers, andDLsequence models\. It also outlines the validation methodology used to assess generalization across heterogeneous deployment conditions\.

### III\-AStatistical Ensemble Method \(SEM\)

SEMprocesses per\-UEKPItelemetry without requiring a training phase\. It adapts to local signal conditions through three component detectors, which are combined using a weighted voting system\[[11](https://arxiv.org/html/2610.20816#bib.bib14),[10](https://arxiv.org/html/2610.20816#bib.bib8)\]\. The parameters are organized into two tiers\. Tier 1 parameters are based on signal\-physics principles and use established default values that do not require labeled data\. This represents the cold\-start mode ofSEM\. Tier 2 denotes supervised refinement, in which selected detector, fusion, and post\-processing parameters are adjusted offline through grid search once labeled data becomes available\. The three detectors, categorized by tier, are detailed below\.

Baseline Statistical Detector:For eachKPIchannelSk​\(t\)S\_\{k\}\(t\), a rolling medianμ^k​\(t\)\\hat\{\\mu\}\_\{k\}\(t\)and rollingMedian Absolute Deviation \(MAD\)σ^k​\(t\)\\hat\{\\sigma\}\_\{k\}\(t\)are computed over a centered window ofWWsamples\. The modified z\-score is\[[11](https://arxiv.org/html/2610.20816#bib.bib14)\]:

zk​\(t\)=0\.6745⋅\|Sk​\(t\)−μ^k​\(t\)\|σ^k​\(t\)\.z\_\{k\}\(t\)=0\.6745\\cdot\\frac\{\|S\_\{k\}\(t\)\-\\hat\{\\mu\}\_\{k\}\(t\)\|\}\{\\hat\{\\sigma\}\_\{k\}\(t\)\}\.\(1\)A sample is flagged ifzk​\(t\)\>τzz\_\{k\}\(t\)\>\\tau\_\{z\}for any channelkk\. The default thresholdτz=3\.0\\tau\_\{z\}=3\.0is used in the Tier 1 cold\-start configuration and corresponds to the standard3​σ3\\sigma\-equivalent rule for outlier detection based on theMAD\. This approach does not require labeled data for validation\. The rolling formulation of the detector reduces sensitivity to long\-term signal drift\. When the data series is too short to allow for adaptive windowing, the method defaults to using global median andMADstatistics\.

Count\-Based Detector:The count ofRSRPsamples, denoted asC⁡\(t\)C\(t\), represents the number ofRSRPsamples successfully acquired during the reporting interval\. Under conditions of high\-gain jamming, this count can drop to zero when the receiver is unable to complete its measurements\. This behavior has been consistently observed across the collected logs\. The detection system flagsC⁡\(t\)=0C\(t\)=0without exception, and it also flags cases where the absolute difference\|C⁡\(t\)−C⁡\(t−1\)\|\|C\(t\)\-C\(t\-1\)\|exceeds an adaptive threshold, which is calculated based on the local rolling standard deviation\. Notably, zero\-count events are not present under normal operating conditions, which contributes to the high precision of this detector\[[10](https://arxiv.org/html/2610.20816#bib.bib8)\]\. In cold\-start operation, the detector uses a default minimum threshold for the adaptive change rule, whereas this threshold can subsequently be refined during supervised grid search\.

Abrupt Change Detector:For each channelkk, we compute the absolute first\-order difference, defined asΔk​\(t\)=\|Sk​\(t\)−Sk​\(t−1\)\|\\Delta\_\{k\}\(t\)=\|S\_\{k\}\(t\)\-S\_\{k\}\(t\-1\)\|\. Samples whereΔk​\(t\)\\Delta\_\{k\}\(t\)exceeds the thresholdTk=\(Skmax−Skmin\)⋅pcT\_\{k\}=\(S\_\{k\}^\{\\max\}\-S\_\{k\}^\{\\min\}\)\\cdot p\_\{c\}are flagged as abrupt change events\. This threshold is self\-normalizing to the signal’s dynamic range, allowing it to adapt automatically to differentKPImagnitudes across deployment sites without requiring labeled data\. The default value forpcp\_\{c\}is0\.40\.4, which is considered a Tier 1 parameter\. Additionally, sustained measurement gaps, where noKPIvalues are reported due to receiver outages, are flagged as implicit anomalies because jammer\-induced receiver disruption can result in systematic measurement loss\[[1](https://arxiv.org/html/2610.20816#bib.bib11)\]\.

Fusion and Decision:The three binary masks,btb\_\{t\}\(baseline\),ctc\_\{t\}\(count\), andata\_\{t\}\(abrupt\), are combined through a weighted vote\[[10](https://arxiv.org/html/2610.20816#bib.bib8)\]:

V⁡\(t\)=wb⋅bt\+wc⋅ct\+wa⋅at,wb\+wc\+wa=1\.V\(t\)=w\_\{b\}\\cdot b\_\{t\}\+w\_\{c\}\\cdot c\_\{t\}\+w\_\{a\}\\cdot a\_\{t\},\\quad w\_\{b\}\+w\_\{c\}\+w\_\{a\}=1\.Three decision masks are derived fromV⁡\(t\)V\(t\)\. The*union mask*flags any sample where at least one detector is activated:Ut=bt∨ct∨atU\_\{t\}=b\_\{t\}\\vee c\_\{t\}\\vee a\_\{t\}\. The*weighted mask*flags samples where the confidence exceeds a thresholdθ\\theta:Wt=1W\_\{t\}=1ifV⁡\(t\)≥θV\(t\)\\geq\\theta, and00otherwise\. The*hybrid mask*is

Ht=ct∨\(bt∧V⁡\(t\)≥0\.5\)∨\(at∧V⁡\(t\)≥0\.6\),H\_\{t\}=c\_\{t\}\\vee\(b\_\{t\}\\wedge V\(t\)\\geq 0\.5\)\\vee\(a\_\{t\}\\wedge V\(t\)\\geq 0\.6\),where∨\\veedenotes logical OR and∧\\wedgedenotes logical AND\. The hybrid rule gives the count detector unconditional priority, requiring the remaining detectors to reach a confidence threshold before contributing\. All reported results utilize the hybrid mask, which consistently outperformed the union and weighted variants across validation conditions by reducing false alarms while preserving recall\. Adjacent detections that are separated by less than a configurable gap are merged, and isolated detections that are shorter than a minimum duration are suppressed using a persistence filter\. The voting weights\{wb,wc,wa\}\\\{w\_\{b\},w\_\{c\},w\_\{a\}\\\}, the gap threshold, the persistence duration, and the thresholdθ\\thetaare Tier 2 parameters\. These parameters have no basis in signal physics and are optimized through a grid search over the available labeled logs\.

### III\-BLearning\-Based Models

We evaluate fiveMLclassifiers:Random Forest \(RF\), eXtreme Gradient Boosting \(XGBoost\), Light Gradient Boosting Machine \(LightGBM\),Support Vector Machine \(SVM\), andLogistic Regression \(LR\)\. We also evaluate fiveDLsequence models:Long Short\-Term Memory \(LSTM\),Gated Recurrent Unit \(GRU\),one\-dimensional Convolutional Neural Network \(1D\-CNN\),Convolutional Neural Network \(CNN\)\-LSTM, andTemporal Convolutional Network \(TCN\)\[[4](https://arxiv.org/html/2610.20816#bib.bib5),[5](https://arxiv.org/html/2610.20816#bib.bib2),[6](https://arxiv.org/html/2610.20816#bib.bib6)\]\.

The learning\-based methods use a common feature\-engineering pipeline applied to the nineKPIchannels\. The representation includes rolling statistics over windows of 5, 10, and 20 samples; lag and lag\-difference features at offsets of 1, 2, 3, and 5 samples; first\- and second\-order changes; volatility and statistical descriptors; cross\-KPIrelationships; and selected frequency\-domain features\. When all nineKPIchannels are available, this produces approximately423423features per sample\. TheMLclassifiers operate directly on these feature vectors, whereas theDLmodels process sequences of1010consecutive feature vectors to capture temporal dependencies\. Feature scaling is fitted using the training data and subsequently applied to validation and test data\.

TheMLclassifiers are trained using their respective classification objectives, with class balancing applied where supported\. TheDLmodels use class\-weighted binary cross\-entropy to address label imbalance, with early stopping based on validation performance\. All fiveMLclassifiers and all fiveDLarchitectures are evaluated in each fold, and the model achieving the highest test\-set F1\-score within each family is reported as the bestMLorDLmodel for that fold\.

### III\-CValidation Methodology

*Single\-environment validation*evaluates each log independently using a temporal split that preserves sample order\. By default,70%70\\,\\%of each log is used for training,15%15\\,\\%for validation, and the remaining15%15\\,\\%for testing\. This setting represents a detector trained and evaluated under the same underlying network conditions\. When necessary, the validation/test boundary is adjusted to ensure that both jamming and non\-jamming samples are represented in the test partition\.

*Multi\-environmentLOOvalidation*holds out one complete log for testing while the remaining seven logs form the training corpus\. Each log is used as the held\-out test environment once\. For theDLmodels, a validation subset is additionally drawn from the multi\-environment training corpus for early stopping\. This regime represents deployment in an environment that was excluded from the training corpus and therefore provides a direct test of generalization across heterogeneous conditions\.

SEMdoes not undergo model training in either validation regime\. In cold\-start mode, it operates using its default parameter configuration without labeled\-data calibration\. In supervised mode, its parameters are refined through grid search using available labeled data\. BecauseSEMis not trained using the single\- or multi\-environment learning corpora, its operation is independent of the training regime used for theMLandDLmodels\. Consequently, the cold\-startSEMresults are identical across the two validation settings\.

## IVExperimental Setup

This section describes the dataset, feature representation, and evaluation metrics used to assess the proposed and learning\-based detection methods under varying deployment conditions\.

### IV\-ADataset

TABLE I:Dataset summary\.The dataset in Table[I](https://arxiv.org/html/2610.20816#S4.T1)includes eight experimental logs collected over approximately2525hours from a5GNRtestbed operating on Band N78 under active jamming conditions\[[3](https://arxiv.org/html/2610.20816#bib.bib1)\]\. The jamming transmission gains were set at4545dB and4848dB\. Network traffic varied from idle to1111Mb/s downlink, withInternet Control Message Protocol \(ICMP\)ping traffic included\. Ground truth labels indicating the actual jamming intervals were manually verified during the experiments, allowing for parameter optimization and performance evaluation\.

The dataset encompasses three axes of variation that serve to stress\-test the methods evaluated in this paper\. The number ofUEs, ranging from 1 to 3, impacts signal diversity and measurement statistics, giving learning\-based models a structural advantage with multi\-UElogs\. The data volume, which ranges from926926to65,07565\{,\}075samples, determines the density of training examples for each deployment condition, directly testing the sample efficiency of each approach\. Lastly, the burst patterns, either fixed at3030or6060seconds or variable between1010and3030seconds, introduce different levels of temporal complexity in the jamming signature\. Models trained on fixed\-burst logs must be able to generalize to variable\-burst conditions during multi\-environment validation, which represents the primary source of distribution shift in this dataset\.

### IV\-BKPIFeatures

NineKPIchannels are extracted from the5Gtestbed measurement reports: the minimum, first quartile \(Q1\), median, third quartile \(Q3\), maximum, and mean ofRSRP;RSRPsample count; downlinkRSSI; and uplinkSNR\. Together, these channels capture three principal effects of broadband jamming on the radio interface: changes in received signal strength, loss of measurement availability, and changes in the observed noise and interference level\.

For the learning\-based methods, the nineKPIchannels are transformed through the feature\-engineering pipeline described in Section[III](https://arxiv.org/html/2610.20816#S3)\. The representation includes rolling statistics over windows of55,1010, and2020samples; lag and lag\-difference features at offsets of11,22,33, and55samples; first\- and second\-order change features; statistical descriptors; cross\-KPIrelationships; and selected frequency\-domain features\. This produces approximately423423features per sample when all nineKPIchannels are available\. TheMLclassifiers operate on individual feature vectors, whereas theDLmodels use sequences of1010consecutive feature vectors\. In contrast,SEMoperates directly on the nine originalKPIchannels without feature engineering\.

### IV\-CEvaluation Metrics

![Refer to caption](https://arxiv.org/html/2610.20816v1/heatmap_single.png)\(a\)Single\-environment validation\.
![Refer to caption](https://arxiv.org/html/2610.20816v1/heatmap_multi.png)\(b\)Multi\-environmentLOOvalidation\.

Fig\. 1:Per\-fold F1\-score, precision, recall, accuracy, and specificity forSEM\(supervised\), the best\-performingMLmodel, and the best\-performingDLmodel under both validation regimes\.We report F1\-score, precision, recall, accuracy, and specificity at the sample level\. Metrics are computed separately for each held\-out log to expose performance variation across deployment conditions, while mean values across the eight logs are used to summarize overall performance\. F1\-score captures the balance between precision and recall, whereas specificity quantifies the ability to correctly identify non\-jamming samples and therefore provides an indication of false\-alarm behavior\. In military jamming detection, recall is particularly important because a missed jamming event can have greater operational consequences than a false alarm\. We therefore interpret recall alongside F1\-score throughout the results rather than relying on aggregate accuracy alone\. For each fold, all fiveMLclassifiers and all fiveDLarchitectures are evaluated, and the highest\-F1 model within each family is reported as the best\-performingMLandDLmodel, respectively\.

## VResults

This section presents the overall performance, per\-log distribution shift analysis, failure mode characterization, and detailed individual model breakdown\.

### V\-AOverall Performance

TABLE II:Mean performance across 8 logs\.SEMcold\-start uses no labeled data;SEMsupervised uses grid\-search\-optimized parameters\. Bold entries highlight the leading value per metric column based on the unrounded mean values\.Table[II](https://arxiv.org/html/2610.20816#S5.T2)presents the average performance across all eight logs forSEMin cold\-start mode,SEMin supervised mode, and the best results from bothMLandDLunder both validation regimes\. The analysis is structured around three key findings\.

Cold\-start parity with recall advantage\.SEM’s cold\-start method achieves a mean F1\-score of0\.8880\.888without labeled data, just0\.0230\.023points lower than single\-environmentMLtrained on70%70\\,\\%of each log\. WhileSEM’s cold\-start precision stands at0\.8260\.826, compared toML’s precision of0\.9840\.984, the difference is largely attributed to a higher false alarm rate\. This is evident fromSEM’s cold\-start recall of0\.9710\.971, in contrast to0\.8900\.890forML\. In military jamming detection, where missed detections are more critical than false alarms,SEM’s cold\-start is the stronger baseline despite its lower F1\-score\.

Grid search recovers F1\-score by\+0\.074\+0\.074, mainly from two specific logs\.SupervisedSEMachieves a mean F1\-score of0\.9620\.962, which remains consistent across both validation regimes\. The largest gains over cold\-start operation occur for Log 6 \(\+0\.195\+0\.195with926926samples, experiencing adaptive windowing instability at short durations\) and Log 8 \(\+0\.151\+0\.151with an1818\-hour low\-gain long duration\)\. Log 3 improves by\+0\.070\+0\.070, while improvements across the remaining five logs range from approximately\+0\.018\+0\.018to\+0\.046\+0\.046, indicating that the Tier 1 default parameters are generally robust across most deployment conditions\.

Corpus coverage drivesMLandDLrecovery\.Multi\-environment training improves theMLF1\-score from0\.9110\.911to0\.9620\.962and theDLF1\-score from0\.8970\.897to0\.9560\.956compared to single\-environment training\. This improvement is primarily due to the diversity of the training data rather than differences in hyperparameter selection\. This finding indicates that corpus coverage is a key factor limiting the performance of learning\-based models when faced with distribution shifts\.

![Refer to caption](https://arxiv.org/html/2610.20816v1/oran_deployment_figure.png)Fig\. 2:SEMdeployment within theO\-RANRICarchitecture and three\-phase deployment sequence\.
### V\-BDistribution Shift Analysis

The overall averages in Table[II](https://arxiv.org/html/2610.20816#S5.T2)mask significant variations across individual logs\. Fig\.[1](https://arxiv.org/html/2610.20816#S4.F1)illustrates the per\-fold performance metrics for all eight logs, revealing four distinct performance patterns\.

1. 1\.Stable high\-gain conditions \(Logs𝟏\\mathbf\{1\}and𝟐\\mathbf\{2\},𝟒𝟖\\mathbf\{48\}dB, multi\-UE, fixed𝟔𝟎\\mathbf\{60\}s\)\.SupervisedSEMand the learning\-based methods achieve F1\-scores of approximately0\.980\.98or higher in both regimes\. The presence of strong interference signatures, a sufficient number of samples, and regular burst patterns allows all methods to effectively detect jamming in these logs, irrespective of the training strategy used\.SEM’s cold\-start method achieves F1\-scores between0\.930\.93and0\.970\.97in this context, even without any labeled data\.
2. 2\.Sparse data and short recordings \(Logs𝟑\\mathbf\{3\},𝟒\\mathbf\{4\}, and𝟔\\mathbf\{6\},𝟗𝟐𝟔\\mathbf\{926\}–5,132\\mathbf\{5\{,\}132\}samples,𝟏𝟓\\mathbf\{15\}–𝟒𝟐\\mathbf\{42\}min\)\.SupervisedSEMmaintains an F1\-score of0\.970\.97\-0\.980\.98across all three logs\. However, the cold\-start performance ofSEMdegrades significantly on Log 6, with an F1\-score of0\.780\.78and precision of0\.690\.69\. In this case, the rolling window ofW=50W=50samples accounts for approximately5%5\\,\\%of the926926\-sample recording, making the adaptive baseline more sensitive to the short recording duration\. In contrast, bothMLandDLapproaches experience moderate degradation in the single\-environment scenario for Log 3 \(ML:0\.900\.90,DL:0\.880\.88\)\. Under multi\-environment training,MLrecovers to0\.970\.97, whileDLreaches0\.910\.91\. Recovery is also not complete for Log 4, where the multi\-environmentMLperformance is0\.900\.90, indicating that the operating conditions of the held\-out log are not fully represented by the remaining training corpus\.
3. 3\.Variable burst pattern, distribution shift regime \(Log𝟕\\mathbf\{7\},\[𝟏𝟎,𝟑𝟎\]\\mathbf\{\[10,30\]\}s bursts,𝟒𝟓\\mathbf\{45\}dB\)\.Log 7 generates the most significant operational results\. In a single\-environment training scenario,MLperformance degrades to an F1\-score of0\.450\.45with a recall of0\.290\.29, failing to detect71%71\\,\\%of the labeled jamming samples\. Meanwhile,DLachieves an F1\-score of0\.640\.64with a recall of0\.470\.47\. In contrast,SEMin cold\-start mode achieves an F1\-score of0\.930\.93with a recall of0\.960\.96across both regimes; this represents a0\.480\.48\-point advantage in F1\-score and a0\.670\.67\-point advantage in recall over single\-environmentMLfor the most challenging distribution\-shift case\. Under multi\-environment training, both learning\-based methods recover to an F1\-score of approximately0\.980\.98, comparable to the0\.970\.97F1\-score of supervisedSEM\. This recovery underscores that diverse training data can address distribution shifts when available, but prior exposure to similar variable\-burst behavior is necessary before deployment\.
4. 4\.Large dataset with underrepresented conditions \(Log𝟖\\mathbf\{8\},𝟏𝟖\\mathbf\{18\}hr,\[𝟏𝟎,𝟑𝟎\]\\mathbf\{\[10,30\]\}s bursts,𝟒𝟓\\mathbf\{45\}dB,65,075\\mathbf\{65\{,\}075\}samples\)\.Log 8 demonstrates that a larger training corpus does not necessarily improve generalization when the target operating condition is insufficiently represented\. The F1\-score forMLdeclines from0\.990\.99in single\-environment validation to0\.890\.89in multi\-environment validation\. Similarly,DLdeclines from0\.980\.98to0\.890\.89, while supervisedSEMremains at approximately0\.880\.88across both regimes\. The training corpus for the multi\-environment fold of Log 8 consists of seven shorter logs that do not adequately represent its1818\-hour low\-gain jammer structure with variable bursts\. This is a dataset coverage limitation for the learning\-based methods, indicating that multi\-environment training is not universally advantageous\. Its effectiveness depends on whether the training data adequately covers the conditions present in the target deployment environment\.

Taken together, Logs 7 and 8 illustrate two complementary aspects of cross\-environment generalization\. Log 7 represents a recoverable distribution shift: learning\-based methods perform poorly under single\-environment training but recover when the training corpus provides broader exposure to heterogeneous operating conditions\. Log 8 presents the opposite case\. Although the multi\-environment training corpus contains substantially more aggregate samples, those samples do not sufficiently represent the long\-duration, low\-gain, variable\-burst behavior of the held\-out environment\. Thus, the two logs demonstrate that training\-set size and training\-set coverage are distinct quantities\. Increasing the amount of labeled data is beneficial only when the resulting corpus captures the conditions likely to occur after deployment\.

### V\-CFailure Mode Analysis

Table[III](https://arxiv.org/html/2610.20816#S5.T3)highlights the key differences betweenSEMand learning\-based methods in the context of Log 7, which presents the most challenging distribution shift\.SEM’s cold\-start mode demonstrates a safe failure mode: while precision decreases, recall remains intact, which means that jamming samples are still detected but with some false alarms\. In contrast, single\-environmentMLandDLfail in an unsafe manner; recall drops to0\.290\.29and0\.470\.47, respectively, resulting in a significant number of undetected jamming samples\. In militaryC2networks, failing to detect jamming activity is an operationally unacceptable outcome\.

TABLE III:Per\-method performance on Log 7\.
### V\-DIndividual Model Analysis

Beyond the best\-per\-family results, the individual model analysis reveals substantial variation within both learning\-based families\. In single\-environment training, LightGBM achieves the highest F1\-score of0\.910\.91amongMLclassifiers, while the1D\-CNNscores0\.880\.88amongDLmodels\. Multi\-environment training substantially boosts performance, withRFreaching an F1\-score of0\.960\.96and XGBoost reaching0\.940\.94\.CNN\-based models, including1D\-CNNandCNN\-LSTM, also score0\.940\.94\. However, two models are notable outliers:SVMscores0\.200\.20, whileTCNscores only0\.100\.10under multi\-environment training\.

## VIO\-RAN Deployment Implications

Fig\.[2](https://arxiv.org/html/2610.20816#S5.F2)illustratesSEM’s deployment within theO\-RANRICarchitecture\.SEM’s ensemble detector runs as anxAppon the near\-RTRIC, receiving liveKPItelemetry via the E2 interface and producing per\-sample jamming decisions in near real time\. The Non\-RTRIChosts the optimizationrApp, which runs grid search over selected detector, fusion, and post\-processing parameters using available labeled logs and delivers the optimized parameters to thexAppthrough the A1 interface\[[2](https://arxiv.org/html/2610.20816#bib.bib12),[3](https://arxiv.org/html/2610.20816#bib.bib1)\]\. This effectively separates real\-time detection from offline optimization within theO\-RANstack\.

Phase 1: Immediate Cold\-Start:SEMis deployed with Tier 1 defaults upon initial activation, allowing detection processing to begin immediately\. Although precision is lower in this mode \(mean0\.8260\.826\), there are elevated false alarm rates in very short recordings and during long\-duration low\-gain conditions\. However, recall remains≥0\.88\\geq 0\.88across all conditions\.

Phase 2: Supervised Refinement:As labeled incidents accumulate, therAppperforms a grid search and delivers optimized parameters through A1 without interrupting live detection\. The improvements are largest for Log 6\-type short recordings, which show a gain of\+0\.195\+0\.195in F1\-score, and for Log 8\-type long\-duration low\-gain conditions, which show a gain of\+0\.151\+0\.151in F1\-score\. Log 3 improves by\+0\.070\+0\.070, while gains across the remaining five logs range from approximately\+0\.018\+0\.018to\+0\.046\+0\.046\.

Phase 3: Optional Transition:Once the dataset sufficiently covers the target deployment conditions, therAppassesses whether a learning\-based alternative has achieved performance parity withSEM, as evaluated on held\-out data\. Multi\-environmentMLachieves a mean F1\-score of0\.9620\.962, essentially matching the supervisedSEMmean F1\-score of0\.9620\.962\.SEMremains available as a fallback option, given that multi\-environment training may not always be beneficial: Log 8 illustrates that when the training dataset does not adequately cover the target deployment conditions, performance can decline regardless of dataset size\.

## VIIConclusion

This paper presentsSEM, a method for detecting jamming inO\-RAN\-native environments\.SEMis evaluated against fiveMLclassifiers and fiveDLsequence models using eight5Glogs\. The main finding is that supervisedSEMoutperforms single\-environment learning\-based methods in mean F1\-score, while cold\-startSEMprovides high recall without requiring labeled data\. While learning\-based models struggle with recall during distribution shifts,SEMtends to preserve detection capability at the cost of additional false alarms\. Multi\-environmentMLachieves a mean F1\-score of0\.9620\.962, essentially matching supervisedSEMat0\.9620\.962, although this similarity in mean performance masks substantial variation across individual logs\. Additionally, performance improvements under multi\-environment training from0\.9110\.911to0\.9620\.962forMLand from0\.8970\.897to0\.9560\.956forDLdemonstrate the importance of diverse training data, indicating that corpus coverage is a key limitation for learning\-based methods in heterogeneous environments\.

## Acknowledgment

This work is partially funded by the Department of Homeland Security’s Science and Technology Directorate \(S&T\)\. U\.S\. Government work, not subject to U\.S\. Copyright\.

## References

- \[1\]H\. Pirayesh and H\. Zeng\(2022\)Jamming attacks and anti\-jamming strategies in wireless networks: A comprehensive survey\.IEEE communications surveys & tutorials24\(2\),pp\. 767–809\.Cited by:[§I](https://arxiv.org/html/2610.20816#S1.p1.1),[§II](https://arxiv.org/html/2610.20816#S2.p1.1),[§III\-A](https://arxiv.org/html/2610.20816#S3.SS1.p4.1)\.
- \[2\]P\. Kryszkiewicz and M\. Hoffmann\(2023\)Open RAN for detection of a jamming attack in a 5G network\.In2023 IEEE 97th Vehicular Technology Conference \(VTC2023\-Spring\),pp\. 1–2\.Cited by:[§I](https://arxiv.org/html/2610.20816#S1.p1.1),[§II](https://arxiv.org/html/2610.20816#S2.p1.1),[§VI](https://arxiv.org/html/2610.20816#S6.p1.1)\.
- \[3\]H\. Bogucka, M\. Hoffmann, P\. Kryszkiewicz, and L\. Kulacz\(2025\)An Open\-RAN Testbed for Detecting and Mitigating Radio\-Access Anomalies\.IEEE Communications Magazine\.Cited by:[§I](https://arxiv.org/html/2610.20816#S1.p1.1),[§II](https://arxiv.org/html/2610.20816#S2.p1.1),[§IV\-A](https://arxiv.org/html/2610.20816#S4.SS1.p1.1),[§VI](https://arxiv.org/html/2610.20816#S6.p1.1)\.
- \[4\]S\. Jere, Y\. Wang, I\. Aryendu, S\. Dayekh, and L\. Liu\(2023\)Bayesian inference\-assisted machine learning for near real\-time jamming detection and classification in 5G New Radio \(NR\)\.IEEE Transactions on Wireless Communications23\(7\),pp\. 7043–7059\.Cited by:[§I](https://arxiv.org/html/2610.20816#S1.p2.1),[§II](https://arxiv.org/html/2610.20816#S2.p2.1),[§III\-B](https://arxiv.org/html/2610.20816#S3.SS2.p1.1)\.
- \[5\]S\. Hong, K\. Kim, and S\. Lee\(2023\)A hybrid jamming detection algorithm for wireless communications: Simultaneous classification of known attacks and detection of unknown attacks\.IEEE Communications Letters27\(7\),pp\. 1769–1773\.Cited by:[§I](https://arxiv.org/html/2610.20816#S1.p2.1),[§II](https://arxiv.org/html/2610.20816#S2.p2.1),[§III\-B](https://arxiv.org/html/2610.20816#S3.SS2.p1.1)\.
- \[6\]M\. Varotto, S\. Valentin, and S\. Tomasin\(2024\)Detecting 5G signal jammers using spectrograms with supervised and unsupervised learning\.In2024 IEEE International Conference on Communications Workshops \(ICC Workshops\),pp\. 767–772\.Cited by:[§I](https://arxiv.org/html/2610.20816#S1.p2.1),[§II](https://arxiv.org/html/2610.20816#S2.p2.1),[§III\-B](https://arxiv.org/html/2610.20816#S3.SS2.p1.1)\.
- \[7\]Y\. Arjoune, F\. Salahdine, M\. S\. Islam, E\. Ghribi, and N\. Kaabouch\(2020\)A novel jamming attacks detection approach based on machine learning for wireless communication\.In2020 International Conference on Information Networking \(ICOIN\),pp\. 459–464\.Cited by:[§I](https://arxiv.org/html/2610.20816#S1.p2.1)\.
- \[8\]S\. Gecgel, C\. Goztepe, and G\. K\. Kurt\(2019\)Jammer detection based on artificial neural networks: A measurement study\.InProceedings of the ACM Workshop on Wireless Security and Machine Learning,pp\. 43–48\.Cited by:[§I](https://arxiv.org/html/2610.20816#S1.p2.1),[§II](https://arxiv.org/html/2610.20816#S2.p2.1)\.
- \[9\]P\. Szilágyi and S\. Nováczki\(2012\)An automatic detection and diagnosis framework for mobile communication systems\.IEEE transactions on Network and Service Management9\(2\),pp\. 184–197\.Cited by:[§II](https://arxiv.org/html/2610.20816#S2.p1.1)\.
- \[10\]G\. F\. Ciocarlie, U\. Lindqvist, S\. Nováczki, and H\. Sanneck\(2013\)Detecting anomalies in cellular networks using an ensemble method\.InProceedings of the 9th international conference on network and service management \(CNSM 2013\),pp\. 171–174\.Cited by:[§II](https://arxiv.org/html/2610.20816#S2.p1.1),[§III\-A](https://arxiv.org/html/2610.20816#S3.SS1.p1.1),[§III\-A](https://arxiv.org/html/2610.20816#S3.SS1.p3.1),[§III\-A](https://arxiv.org/html/2610.20816#S3.SS1.p5.1)\.
- \[11\]M\. Cheng, Y\. Ling, and W\. B\. Wu\(2017\)Time series analysis for jamming attack detection in wireless networks\.InGLOBECOM 2017\-2017 IEEE Global communications conference,pp\. 1–7\.Cited by:[§II](https://arxiv.org/html/2610.20816#S2.p1.1),[§III\-A](https://arxiv.org/html/2610.20816#S3.SS1.p1.1),[§III\-A](https://arxiv.org/html/2610.20816#S3.SS1.p2.1)\.
- \[12\]S\. Wang, H\. Wang, W\. Wang, and V\. C\. Leung\(2023\)Detecting intelligent jamming on physical broadcast channel in 5G NR\.IEEE Communications Letters27\(5\),pp\. 1292–1296\.Cited by:[§II](https://arxiv.org/html/2610.20816#S2.p1.1)\.
- \[13\]L\. Chiarello, P\. Baracca, K\. Upadhya, S\. R\. Khosravirad, and T\. Wild\(2021\)Jamming detection with subcarrier blanking for 5G and beyond in industry 4\.0 scenarios\.In2021 IEEE 32nd Annual International Symposium on Personal, Indoor and Mobile Radio Communications \(PIMRC\),pp\. 758–764\.Cited by:[§II](https://arxiv.org/html/2610.20816#S2.p1.1)\.
- \[14\]M\. Hoffmann and P\. Kryszkiewicz\(2023\)Signaling storm detection in IIoT network based on the open RAN architecture\.InIEEE INFOCOM 2023\-IEEE Conference on Computer Communications Workshops \(INFOCOM WKSHPS\),pp\. 1–2\.Cited by:[§II](https://arxiv.org/html/2610.20816#S2.p1.1)\.