Source-linked AI summary
Diffusion-Based Stress Testing of Overload Monitoring for Resilient Emergency Cellular Networks Using Internet CDR Proxies
Bilal Hussain, Xiao Tang, Tan Li, Muhammad Azhar, Danista Khan, Fawad Ahmad
TL;DR
Fine-grained cellular signaling telemetry is costly and privacy-sensitive, so the paper investigates Internet activity in coarse CDR grids as a proxy for hidden overload stress. It trains a lightweight detector on stylized surges, stress-tests it with diffusion-synthesized surges, and adapts it using hard synthetic samples. The stress test reveals default-threshold failure despite useful ranking, while adaptation restores thresholded performance, subject to oracle-threshold and seed-variance caveats.
Problem
Fine-grained RRC or NGAP telemetry is privacy-sensitive and costly to collect, while emergency monitoring pipelines often rely on coarse CDR aggregates that do not directly measure signaling overload.
Source: §I · p. 1Method
The study treats Internet CDR activity as an indirect overload proxy, trains a lightweight detector on stylized surges, stress-tests it with diffusion-synthesized events, and adapts it on hard synthetic samples.
Source: §I · p. 1Results
At τ = 0.5, internet_only F1 falls to 0.00 ± 0.00% under stress testing despite ROC-AUC of 0.886 ± 0.040, then reaches 85.67 ± 14.37% F1 and 0.99996 ± 0.00007 ROC-AUC after adaptation.
Source: A. Baseline and Stress-Test Behavior · p. 4, D. Recovery After Adaptation · p. 5, §V · p. 6Takeaways & Limitations
The diffusion protocol provides a reusable test platform for disaster-like surge shift, and operating-point drift can inform recalibration before overload actions.
Source: §V · p. 6, §VI · p. 6The post-adaptation F1 varies substantially across three seeds, and oracle τ* values are selected post hoc on the evaluation grids rather than independently validated operating points.
Source: §V · p. 6
Abstract
from arXiv · showhide
Disasters can overload cellular control-plane signaling within minutes, yet fine-grained Radio Resource Control (RRC) or Next Generation (NG) Application Protocol (NGAP) telemetry is privacy-sensitive and costly to collect for analytics. Many emergency monitoring pipelines therefore rely on coarse Call Detail Record (CDR) aggregates. We treat Internet activity in CDR grids as a practical proxy for hidden signaling stress under that constraint. We train a lightweight convolutional neural network (CNN) on stylized overload injections, stress-test it with diffusion-synthesized surges that preserve normal traffic structure, and adapt the detector by retraining on hard synthetic samples. Under stress-test conditions, the default alert threshold fails even though receiver operating characteristic (ROC) curves stay strong: the detector still assigns overloaded cells a larger overload probability than normal cells, but those probabilities fall below the default cutoff 0.5 and are labeled normal, so the F1-maximizing threshold -- selected post hoc on the same stress-test grids (oracle $\tau^*$) -- shifts by $0.32 \pm 0.03$ (operating-point drift). Across three random seeds, hard-sample adaptation raises thresholded performance (F1) from 0% (no alerts at the default cutoff 0.5 on any seed) to $85.67 \pm 14.37$% and ranking from ROC-AUC $0.886 \pm 0.040$ to $0.99996 \pm 0.00007$. Diffusion-synthesized surges expose threshold fragility that matched-condition training -- training and testing on the same stylized injections -- hides, and hard-sample adaptation restores usable alerts at the default cutoff. Together, these steps define a reusable pre-deployment stress test for emergency monitors. Internet-only CDR input further supports lightweight AI-native workflows that combine monitoring, recalibration, and adaptation.
I. INTRODUCTION
Disaster surges can stress cellular control-plane signaling, while fine-grained telemetry is costly and privacy-sensitive, motivating coarse CDR-based monitoring. This study uses Internet CDR activity as an indirect proxy and tests a diffusion-based stress protocol with adaptation and threshold-drift analysis.
I. INTRODUCTION
- Disaster-related control-plane overload can degrade cellular service within minutes, making rapid coarse-grained stress indicators relevant to emergency response.The introduction cites disasters and an urban fire as motivating examples for dependable alerting and overload monitoring.Source: §I · p. 1
- Internet CDR activity is treated as an indirect proxy for hidden control-plane stress because fine-grained RRC or NGAP telemetry is privacy-sensitive and costly to collect.The Milano CDR collection supplies real cell-level grids for learning normal behavior and constructing labeled synthetic surges.Source: §I · p. 1
- The study emphasizes emergency overload monitoring under distribution shift, explicit threshold-drift analysis, and post-shift adaptation rather than channel-aligned CDR detection.It complements a companion TII study by focusing on Internet-only emergency overload, multi-seed thresholded F1, and calibration-only versus hard-sample adaptation.Source: §I · p. 1
- Diffusion-synthesized surges expose threshold fragility under distribution shift, while hard-sample adaptation and operating-point analysis assess recovery.The work frames the protocol as a reusable evaluation platform for emergency monitors.Source: §I · p. 2
- Internet-only compact-CNN monitoring is compared with a three-channel ResNet50 input–architecture pair on the proxy task.The stated motivation is a lightweight monitoring workflow using coarse CDR aggregates.Source: §I · p. 2
II. RELATED WORK
Prior work uses CDR aggregates for cellular analytics and diffusion models for realistic synthetic traffic, but this paper centers their use on emergency overload monitoring under shift. It evaluates an Internet-proxy detector, threshold drift, adaptation, and diffusion-surge realism.
II. RELATED WORK
- CDR aggregates support anomaly detection, forecasting, and cell-level analytics, while public corpora such as Milano enable reproducible studies.Earlier work commonly aligns the monitored CDR channel with the event, such as SMS counts for SMS flooding or Internet counts for data overload.Source: §II · p. 2
- Prior CDR-grid studies often monitor the channel directly reflecting the event, whereas this study uses Internet activity as a proxy for control-plane stress under distribution shift.The paper compares its focus with channel-aligned detectors and evaluates operating-point drift, adaptation, and diffusion-surge realism.Source: §II · p. 2
- Diffusion models generate traffic samples that preserve aggregate structure and can support harder out-of-distribution samples, while matched stylized train–test evaluation can overstate reliability.This paper uses diffusion as an emergency-network evaluation and threshold-calibration probe, rather than only as data augmentation.Source: §II · p. 2
III. METHODOLOGY
The methodology evaluates cell-level overload monitoring through matched stylized surges, diffusion-synthesized stress tests, and adaptation on hard synthetic samples. Its main detector processes Internet-only 9 × 9 CDR snapshots with a compact CNN, with online thresholding separated from offline evaluation and adaptation.
III. METHODOLOGY
- The protocol has three stages: matched stylized-surge baseline evaluation, diffusion-synthesized stress testing, and adaptation after retraining on hard synthetic samples.Multi-seed runs vary initialization, shuffling, and the selection of overload-affected cells.Source: §III · p. 2
A. Problem Setting and System Model
- Each observation is a 9 × 9 × C grid snapshot, and the main Internet-only detector uses a compact CNN with three convolution–pooling blocks and an 81-unit sigmoid output.A three-channel comparison uses ResNet50; these settings differ in both input and architecture.Source: A. Problem Setting and System Model · pp. 2–3
- The online path scores 9 × 9 snapshots and thresholds at fixed τ = 0.5, while diffusion stress testing, adaptation, and oracle τ* selection occur offline.The offline τ* is selected on stress-test grids and is not the online cutoff.Source: A. Problem Setting and System Model · p. 3
B. Proxy Validity Rationale
The study treats Internet CDR activity as an indirect indicator of control-plane stress, not a direct signaling measurement. It uses Milano traffic and synthetically labels overload cells, retaining the corpus as a reproducible benchmark despite its 3G/4G-era scope.
B. Proxy Validity Rationale
- Internet activity in CDR aggregates can co-occur with control-plane overload during disasters, but it does not measure signaling directly.The evaluation asks whether the proxy preserves relative overload probabilities under shift and whether recalibration or adaptation restores usable alerts, without claiming one-to-one equivalence to RRC load.Source: B. Proxy Validity Rationale · p. 3
C. Dataset Scope, Grid Construction, and Baseline Surges
- The Milano corpus covers 2013–2014 metropolitan 3G/4G-era traffic rather than a 5G standalone core capture.The study does not claim that its decade-old traffic statistics match modern 5G loads.Source: C. Dataset Scope, Grid Construction,… · p. 3
- Milano is retained as a reproducibility benchmark because no public labeled per-cell cellular-overload CDR corpus offers a drop-in replacement.The study learns normal traffic from real CDR data and synthesizes labeled surge grids to evaluate distribution shift.Source: C. Dataset Scope, Grid Construction,… · p. 3
- Overload is defined by a synthesized Internet surge, not measured RRC; α = 43.3 follows the Internet-signaling emulation, and the surge replaces the original Internet count in cell k [12].Normal grids are unperturbed Milano snapshots. The stylized surge is approximately 59σ separated from normal Internet aggregates.Source: C. Dataset Scope, Grid Construction,… · p. 3
D. Diffusion Stress-Test Generator
The stress-test generator learns the structure of normal three-channel CDR aggregates, then creates near-manifold surge samples by perturbing Internet activity. These samples are used for stress-test evaluation rather than as paired disaster traces.
D. Diffusion Stress-Test Generator
- The diffusion model is fitted to normal three-channel CDR aggregates and generates rows on their learned traffic manifold.The input is x0 = [xSMS, xCall, xInternet] in R^3; the forward process adds noise, and a learned reverse process samples rows.Source: D. Diffusion Stress-Test Generator · p. 3
- The denoiser is trained with a noise-prediction objective on the noised CDR aggregates.The noised input combines the clean row with Gaussian noise according to the cumulative noise-schedule coefficient.Source: D. Diffusion Stress-Test Generator · p. 3
- The generator perturbs only Internet activity with a milder random multiplier to create synthetic surge rows.The perturbations are approximately 3.6σ, compared with the fixed baseline multiplier α=43.3 (approximately 59σ).Source: D. Diffusion Stress-Test Generator · pp. 3–4
- The milder Internet shifts serve as near-manifold, emergency-like stress tests, not paired disaster traces.This procedure is used only during stress-test evaluation.Source: D. Diffusion Stress-Test Generator · p. 4
E. Adaptation with Hard Synthetic Samples
Hard-sample adaptation uses detector feedback to steer diffusion sampling toward difficult overload-like examples, adds them to training, and fine-tunes the detector. The accompanying threshold metric diagnoses shifts in the alert operating point, but its F1-maximizing threshold is an oracle value selected on the evaluation grids.
E. Adaptation with Hard Synthetic Samples
- Detector feedback steers diffusion sampling toward hard, overload-like examples, which are added to training before brief detector fine-tuning.The procedure is a lightweight variant of guided retraining intended to stabilize monitoring under shifted conditions.Source: E. Adaptation with Hard Synthetic… · p. 4
F. Threshold and Resilience Metrics
- High ROC-AUC can coexist with collapsed F1 when overloaded-cell scores remain below the fixed 0.5 cutoff, causing every cell to be labeled normal.The paper reports both ROC-AUC and threshold-dependent F1 after sweeping thresholds from 0.05 to 0.95.Source: F. Threshold and Resilience Metrics · p. 4
- The F1-maximizing threshold and operating-point drift are post-hoc oracle values on the evaluation grids, not independently validated operating points.A large drift indicates that the default alert rule has become fragile under surge shift.Source: F. Threshold and Resilience Metrics · p. 4
G. Evaluation Protocol
The evaluation protocol compares baseline, diffusion stress-test, and adaptation stages across multiple seeds. Reported results are averaged over three seeds, and the protocol’s table and figure summarize the stages and F1 evaluation.
G. Evaluation Protocol
- The protocol evaluates matched stylized surges at baseline, diffusion-synthesized events at stress-test, and the detector after hard-sample retraining at adaptation.Multi-seed runs vary initialization, shuffling, and overload-affected cell selection.Source: G. Evaluation Protocol · p. 4
- Table II summarizes multi-seed evaluation across the three stages: baseline, stress-test, and adaptation.The supplied caption identifies the table's scope but does not specify cell values or a stage comparison outcome.Source: Table II · p. 4
IV. RESULTS
- Reported results average three seeds, while Fig. 2 presents multi-seed F1 by baseline, stress-test, and adaptation stage with one-standard-deviation error bars.Snapshot counts and splits follow Table I.Source: Figure 2 · p. 4, §IV · p. 4
A. Baseline and Stress-Test Behavior
Diffusion-synthesized stress tests expose failures hidden by matched-condition validation: the Internet-only detector’s F1 collapses at the default cutoff despite retained ranking ability. The surge samples also depart less from normal traffic references than multiplier-based injections, though fine-grained spatial realism remains unresolved.
A. Baseline and Stress-Test Behavior
- Under matched baseline conditions, internet_only reaches 99.94 ± 0.05% F1, while three_channel reaches 49.21 ± 10.36%.The stylized surge scales only Internet activity, leaving SMS and call channels near-normal as distractors.Source: A. Baseline and Stress-Test Behavior · p. 4
- At τ = 0.5, internet_only F1 falls to 0.00 ± 0.00% while ROC-AUC remains 0.886 ± 0.040; its oracle threshold shifts to 0.18 ± 0.03 (∆τ = 0.32 ± 0.03).The stress-test results show why a strong ranking score does not guarantee alerts at the default cutoff.Source: A. Baseline and Stress-Test Behavior · p. 4
- Three_channel has stress-test ROC-AUC 0.264 ± 0.033 and ∆τ = 0.45; stress-test F1 is 0.00 for both modes at τ = 0.5.The modes differ in both input and backbone, so the comparison does not isolate a channel effect.Source: A. Baseline and Stress-Test Behavior · p. 4, B. Diffusion Sample Realism · p. 5
B. Diffusion Sample Realism
- TabDDPM stress samples deviate from normal references by 3.61σ (joint MMD2 = 0.877), versus 58.76σ for α = 43.3 multiplier samples.The comparison uses 5,000 normal cell-level references; the authors connect the multiplier deviation to matched-condition over-optimism.Source: B. Diffusion Sample Realism · p. 5
- Internet-channel patterns have weak correlation with the mean normal spatial template: 0.219 ± 0.146 for TabDDPM and 0.248 ± 0.117 for multipliers.Fine-grained spatial realism remains future work.Source: B. Diffusion Sample Realism · p. 5
C. Input Representation Comparison
The compact-CNN internet_only pair outperforms the ResNet50 three_channel pair on stress-relevant metrics, although the comparison changes both input and backbone. The Internet-only design also supports lower-compute monitoring when only coarse CDR summaries are available.
C. Input Representation Comparison
- The compact-CNN internet_only pair has higher stress-test AUC (0.886 vs. 0.264), higher post-adaptation F1 (85.67 vs. 66.80), and smaller stress-test drift (∆τ = 0.317 vs. 0.450).Stress-test F1 is 0.00 for both modes at τ = 0.5; because both input and backbone change, this does not isolate a pure channel effect.Source: C. Input Representation Comparison · p. 5
- The reported τ ∗ is the oracle F1-maximizing threshold selected on the same stress-test grids, with no retraining for rows 1–2.Source: C. Input Representation Comparison · p. 5
- When only coarse CDR summaries can run locally, monitoring the Internet channel alone reduces compute load while yielding clearer overload signatures for this proxy task.Source: C. Input Representation Comparison · p. 5
D. Recovery After Adaptation
Hard-sample adaptation outperforms threshold-only recalibration under stress-test shift, improving both thresholded performance and ranking. The recovery supports a workflow of detecting drift under stress tests and then retraining or recalibrating before a subsequent incident window.
D. Recovery After Adaptation
- After adaptation, internet_only reaches 85.67 ± 14.37% F1 and 0.99996 ± 0.00007 ROC-AUC; its oracle threshold is 0.33 ± 0.06 (∆τ = 0.17).The F1 variation across three seeds is large, so τ = 0.5 F1 is not a certified operating point.Source: D. Recovery After Adaptation · p. 5
- Fig. 4 shows restored strong separability and a narrower gap between ranking and decision quality; the suggested emergency workflow detects drift, then retrains or recalibrates before the next incident window.Source: D. Recovery After Adaptation · p. 5
- The three-channel model reaches 100% F1 on stress-only samples after adaptation but 66.80 ± 0.00% on the mixed normal-plus-stress set.Under this protocol, the compact Internet-only CNN is the stronger of the two input–architecture pairs.Source: D. Recovery After Adaptation · p. 5
E. Calibration-Only Baseline vs. Adaptation
- Threshold-only recalibration raises F1 from 0.00 ± 0.00% to 67.07 ± 5.48%, while hard-sample adaptation reaches 85.67 ± 14.37% at τ = 0.5 and 98.95 ± 1.19% with post-adaptation τF1max.Adaptation improves ranking and thresholded performance beyond calibration alone.Source: E. Calibration-Only Baseline vs.… · p. 5
F. Spatial Consistency
Adaptation largely restores usable alerts at the default threshold, and cell-level accuracy becomes more spatially uniform. These findings are bounded by oracle thresholds, synthetic data from one Milano window, and remaining seed variance.
F. Spatial Consistency
- After adaptation, accuracy becomes more uniform across the 9 × 9 grid, whereas pre-adaptation performance is spatially uneven.The passage connects spatial uniformity to overload actions that should not depend on a handful of easy cells.Source: F. Spatial Consistency · p. 6
V. DISCUSSION AND LIMITATIONS
- Diffusion-synthesized surges expose threshold fragility while preserving useful ranking quality; adaptation largely restores usable alerts at τ = 0.5, with remaining seed variance.Source: §V · p. 6
- Operating-point drift and calibration-versus-adaptation inform planning by showing when fixed alert rules must be recalibrated before overload actions.The diffusion protocol is presented as a reusable testing platform under disaster-like surge shift.Source: §V · p. 6
- Internet CDR is an indirect proxy without paired RRC/NGAP traces, and labels and surges are synthetic on one Milano window.The reported τ* and Δτ are oracle values on evaluation grids; three seeds leave post-adaptation F1 standard deviation at 14.37%.Source: §V · p. 6
VI. CONCLUSION
- Guided adaptation restores usable cell-level alerts at τ = 0.5, while fixed-threshold monitoring can fail under realistic surge shift despite strong ranking quality.The conclusion notes oracle-τ* and seed-variance caveats.Source: §VI · p. 6