PaperlayerView exact revision on arXiv

Source-linked AI summary

Diff-DDoS: Realistic Cyber-Physical Attack Synthesis and Robust Detection for 5G-Enabled CPS Using Tabular Diffusion Models

Bilal Hussain, Xiao Tang, Qinghe Du, Tan Li, Muhammad Azhar, Danista Khan

arXiv:2608.17796v1cs.CRcs.LG

TL;DR

5G-enabled cyber-physical DDoS detectors lack labeled attacks and realistic synthetic data, limiting robustness against adaptive, manifold-preserving attacks. Diff-DDoS uses tabular diffusion and adversarial diffusion training to synthesize realistic attacks and harden detectors, recovering strong F1 across multiple scenarios, including 100% on Internet-signaling attacks.

  • Problem

    5G-enabled cyber-physical DDoS detection lacks scarce labeled attacks and realistic synthetic data that preserve normal traffic semantics.

  • Method

    Diff-DDoS uses TabDDPM to generate manifold-preserving attacks and inverse-guided adversarial diffusion training to iteratively harden detectors.

  • Results

    ResNet50 with ADT recovered F1-scores of 79.6% for silent-call, 100% for Internet, and 92.8% for blended attacks.

  • Takeaways & Limitations

    Tabular diffusion models support realistic stress-testing and robustification of intrusion detectors in data-scarce, 5G-motivated CPS monitoring.

  • Takeaways & Limitations

    External validity is bounded by synthetic evaluation, coarse realism proxies, 3G/4G-era Milano traffic, and ten-minute monitoring unsuitable for sub-millisecond uRLLC control.

Abstract

from arXiv · show

Deep learning-based DDoS detectors for 5G-enabled cyber-physical systems face scarce labeled attack data and unrealistic synthetic substitutes, which limit robustness against adaptive adversaries. Detectors trained on hand-crafted attacks with fixed scaling multipliers degrade catastrophically (F1-score drops of about 47 percent to 100 percent, depending on scenario) when confronted with realistic, distribution-preserving samples. We propose Diff-DDoS, a three-phase framework for realistic attack synthesis and robust detection using tabular diffusion models. Phase 1 trains a baseline CNN cell-level detector on spatiotemporal grids from call detail records (CDRs). Phase 2 trains a tabular denoising diffusion probabilistic model (TabDDPM) on normal CDR aggregates to generate realistic attacks and expose detector vulnerabilities. Phase 3 introduces adversarial diffusion training (ADT), using inverse classifier guidance to generate hard yet distribution-preserving samples until the detector converges. On a Milano CDR dataset across SMS-flooding, silent-call, Internet-signaling, and blended scenarios, ResNet50 with ADT recovers F1-scores of 79.62 percent (silent-call), 100 percent (Internet), and 92.79 percent (blended). After validation-based threshold calibration, ADT reaches 100 percent SMS F1 versus 47.3 percent for CTGAN, and matches the strongest gradient-based adversarial-training baseline on silent-call. These results support tabular diffusion models for stress-testing and hardening intrusion detectors in data-scarce 5G cyber-physical deployments.

I. INTRODUCTION … A. CDR Grid Representation and Detection Task

Diff-DDoS addresses the brittleness of 5G-enabled CPS DDoS detectors by using tabular diffusion to generate realistic, manifold-preserving attacks and adversarially harden cell-level CNNs. The paper builds on CDR grid representations, formalizes multilabel cell detection, and evaluates attacks spanning SMS flooding, silent-call, Internet-signaling, and blended scenarios.

  • I. INTRODUCTION: The study positions robust detection as a necessary response to the expanded attack surface created by 5G integration with industrial monitoring, transportation, and energy CPS services.The introduction motivates multi-layer AI-enabled protection for threats affecting network availability and physical processes.
  • I. INTRODUCTION: Detectors trained on stylized multiplier attacks can fail under realistic distribution-preserving attacks because scarce labels and naive perturbations make near-manifold behavior difficult to model.Silent-call, SMS-flooding, and Internet-signaling attacks stress cellular control-plane procedures rather than only consuming user-plane bandwidth.
  • I. INTRODUCTION: Diff-DDoS uses a three-phase pipeline: a multiplier-based CNN baseline, TabDDPM attack synthesis from normal CDR aggregates, and inverse-classifier-guided adversarial diffusion training.The diffusion-generated samples expose poor baseline generalization while ADT creates hard, realistic attacks for iterative robust training.
  • II. RELATED WORK: CDR-based DDoS research has established strong CNN performance on synthetic multiplier attacks, but related detection and analytics studies rarely test adaptive evasion near normal traffic statistics.Prior work also includes recurrent, federated, and quantum-inspired detectors, while CTGAN augmentation has instability and gradient-based perturbations can be off-manifold.
  • A. CDR Grid Representation and Detection Task: Each input is a 9 × 9 × 3 grid representing SMS-out, Call-out, and Internet aggregates for 81 Milano cells over one 10-minute interval.The selected interval lies within the 11:00–14:00 window.
  • A. CDR Grid Representation and Detection Task: The detection task is multilabel cell-level classification, with a CNN producing sigmoid attack probabilities for each cell against binary cell labels.During ADT, detector parameters are updated iteratively by attack scenario and training round.
  • A. CDR Grid Representation and Detection Task: The attack generator modifies 40 randomly selected cells using scenario-specific multipliers and evaluates SMS flooding, silent-call, Internet-signaling, and blended attacks.These scenarios respectively elevate SMS, suppress calls, elevate Internet traffic, or combine SMS/Internet elevation with call suppression.

B. Denoising Diffusion and Classifier Guidance … A. Data Representations and Attack-Steering Mechanisms

Diff-DDoS combines TabDDPM-based attack synthesis with classifier-guided, inverse-guided diffusion to expose detector weaknesses and iteratively harden detection. Its methodology maps standardized 3D CDR aggregates onto spatiotemporal grids while separating offline synthesis from online inference.

  • B. Denoising Diffusion and Classifier Guidance: TabDDPM applies DDPM forward corruption and learned reverse denoising to 3D CDR aggregates, extending diffusion modeling from images to tabular data.The forward process corrupts clean x_0 into x_t over T steps, while the denoiser predicts injected noise.
  • C. Notation Summary: The main text overloads notation for compact presentation, with Table I summarizing symbols and the supplement providing extended disambiguation.The notation summary distinguishes the symbols used throughout the methodology.
  • B. Denoising Diffusion and Classifier Guidance: Classifier guidance adds a strength-controlled classifier gradient to the unconditional score, biasing reverse diffusion toward a target class.The conditioning strength is denoted s_cg, and the gradient is taken with respect to the noisy sample x_t.
  • B. Denoising Diffusion and Classifier Guidance: Phase 3 reverses classifier guidance, steering generated samples away from attack-like detector scores while preserving the normal CDR manifold.This inverse guidance underpins adversarial diffusion training and builds on prior guidance methods,.
  • IV. METHODOLOGY: Diff-DDoS uses a three-phase pipeline: multiplier-trained CNN baseline, TabDDPM vulnerability exposure, and ADT with inverse guidance for detector hardening.Offline attack synthesis is separated from online grid-level inference.
  • A. Data Representations and Attack-Steering Mechanisms: The methodology represents each normal training row as standardized SMS, call, and Internet aggregates extracted from a 9 × 9 CDR subgrid.StandardScaler is fit only on normal rows, and cell indices map to row-major grid coordinates.
  • A. Data Representations and Attack-Steering Mechanisms: MapToGrid copies a background grid, writes generated aggregates into attacked cells, and labels those cells as attacks; Phase 2 uses abnormal templates, whereas Phase 3 uses zero backgrounds.These representations support the phase-specific attack-steering mechanisms summarized in Table II.

B. Phase 1: Baseline Detector Training · C. Phase 2: TabDDPM-Based Vulnerability Demonstration

Phase 1 trains multilabel CNN detectors on normal and fixed-multiplier attack grids, while Phase 2 replaces test attacks with TabDDPM-generated grids to expose vulnerability to realistic diffusion-based attacks. TabDDPM models 3D CDR aggregates and maps sampled aggregates back to spatial grids, preserving scenario-specific channel control.

  • B. Phase 1: Baseline Detector Training: Phase 1 trains an 81-output multilabel detector on normal grids and multiplier-attack grids with 40 perturbed positive cells, without TabDDPM-generated attacks.Attack labels arise solely from synthesis rather than statistical thresholding.
  • B. Phase 1: Baseline Detector Training: Phase 1 constructs training data through grid extraction and fixed-multiplier injection on |C| = 40 cells, defining labels solely by the synthesis process.The procedure follows Supplementary Algorithms S1–S2 and Phase 1 follows.
  • B. Phase 1: Baseline Detector Training: The baseline compares SimpleCNN with three convolution–pooling blocks against a 50-layer ResNet50 adapted to 9 × 9 × 3 inputs and 81 sigmoid outputs.Both architectures use multilabel binary crossentropy averaged over training images and cell decisions.
  • B. Phase 1: Baseline Detector Training: Predictions are thresholded at τ = 0.5, and TP, FP, and FN are micro-averaged across Ntest×81 cell decisions to compute detection metrics.Reported metrics include Accuracy, Precision, Recall, F1, FPR, and ROC–AUC.
  • C. Phase 2: TabDDPM-Based Vulnerability Demonstration: Phase 2 evaluates multiplier-trained detectors against 168 TabDDPM-generated test attacks combined with 167 held-out normal grids, demonstrating vulnerability to realistic diffusion-based attacks.The working pool contains 335 abnormal-template grids, with multiplier-based training attacks and diffusion-generated test attacks.
  • C. Phase 2: TabDDPM-Based Vulnerability Demonstration: TabDDPM synthesizes attacks from 3D SMS, CALL, and INTERNET CDR aggregates, then MapToGrid(·, B) converts sampled aggregates into 9 × 9 × 3 grids.Modeling the low-dimensional aggregate manifold avoids coupling unrelated cells and supports scenario-specific constraints and per-channel control.

1) Tabular Diffusion Model for Aggregated CDRs: · 2) Realistic Attack Scenario Generation: · D. Phase 3: Adversarial Diffusion Training (ADT)

Diff-DDoS adapts DDPMs to three-dimensional aggregated CDR features, generates scenario-specific attacks through normal-data modeling and mild steering, and hardens detectors with iterative inverse-guided adversarial diffusion training.

  • 1) Tabular Diffusion Model for Aggregated CDRs:: The TabDDPM models the normal CDR manifold using aggregated SMS, CALL, and INTERNET totals for each cell, timeslot, and day.It uses a fixed 1,000-step linear noise schedule and an MLP denoiser with hidden dimensions [256, 512, 256].
  • 1) Tabular Diffusion Model for Aggregated CDRs:: TabDDPM is trained exclusively on normal aggregate CDR rows before being used to synthesize scenario-specific attacks.The model operates on scaled normal aggregates and supports subsequent attack generation.
  • 2) Realistic Attack Scenario Generation:: Reverse diffusion produces aggregate samples, while lightweight heuristic steering periodically targets the scenario’s relevant channel during sampling.This steering occurs every max(1, ⌊T/10⌋) timesteps and is distinct from Phase 3 inverse classifier guidance.
  • 2) Realistic Attack Scenario Generation:: Scenario-specific post-sample constraints apply mild random transformations to target channels: SMS and Internet use 2–5× scaling, while silent-call reduces CALL to 0.3–0.7×.The supplied passage also identifies a blended scenario, but its transformation is truncated in the excerpt.
  • D. Phase 3: Adversarial Diffusion Training (ADT): ADT iteratively appends inverse-guided TabDDPM samples to the training set and fine-tunes the detector for each SMS, Call, Internet, and Blended scenario.The procedure repeats across rounds r = 1, . . . , R using adversarial grids generated in each round.
  • D. Phase 3: Adversarial Diffusion Training (ADT): The evaluation construction uses 167 held-out normal grids and 168 TabDDPM attack grids, yielding 335 test images per scenario.The algorithm records scenario-level F1 without retraining on the assembled test set.
  • D. Phase 3: Adversarial Diffusion Training (ADT): Inverse classifier guidance steers TabDDPM sampling away from attack-like detector scores during each ADT round, producing harder training examples.The resulting adversarial grids are incorporated into detector fine-tuning rather than merely evaluated without retraining.

1) Inverse Classifier Guidance:

Inverse classifier guidance steers TabDDPM reverse diffusion by optimizing the detector’s mean attack probability across 40 replicated cells. ADT applies this guidance iteratively, with channel-specific gradient scaling in non-blended scenarios to avoid distorting unrelated traffic dimensions.

  • Guidance formulation: Inverse guidance computes attack probability as the mean detector score across 40 cells after replicating each CDR aggregate onto a grid.The grid background is drawn from normal training grids during ADT sampling.
  • Guidance formulation: Guidance updates the diffusion sample using the gradient of attack probability toward target probability ptarget = 0.3, scaled by λ.The update is applied during reverse diffusion using the specified guidance scale.
  • Channel protection: For non-blended scenarios, ADT scales gradients in non-target CDR channels by 0.7 before applying inverse guidance.This restricts guidance from altering unrelated traffic dimensions.
  • Iterative ADT procedure: ADT repeatedly generates adversarial samples through TabDDPM reverse diffusion and applies inverse guidance using the detector from the preceding training round.The procedure operates across attack scenarios and updates the detector iteratively after adversarial sample generation.

2) ADT Iterative Training Procedure:

ADT uses inverse guidance with randomly drawn normal training backgrounds, while appended training grids use zero backgrounds and write aggregates only into C. Its hyperparameters are specified in Sec. V.

  • ADT iterative training procedure: Inverse guidance forms I_grid(x) with a randomly drawn normal training background.
  • ADT iterative training procedure: ADT-appended training grids use a zero background and write the aggregate only into C.
  • ADT iterative training procedure: The procedure’s hyperparameters R, N_ADT, E_ADT, λ, and p_target are given in Sec. V.

V. EXPERIMENTAL SETUP

The experiments use Telecom Italia Milano CDR data arranged as spatiotemporal grid images across four attack scenarios, with diffusion, detector, and adversarial-training configurations specified. Evaluation uses standardized aggregates and multiple classification metrics under defined train/test protocols.

  • Dataset: The Milano CDR collection forms 1,116 9 × 9 grid images per scenario across 18 daily time slots, split into 781 training and 335 test images.Training contains 614 normal and 167 attack images; testing contains 167 normal and 168 attack images. Random seed 1 and |C| = 40 govern shuffling and attack-cell selection.
  • Model training: TabDDPM trains for 100 epochs on normal 3D aggregate CDR features, while SimpleCNN and ResNet50 train for up to 100 epochs in Phases 1–2.ResNet50 uses a 10% validation split with early stopping, whereas SimpleCNN uses the full 100-epoch budget.
  • Adversarial diffusion training: ADT uses three rounds with λ = 1.0, 100 generated samples per round and scenario, 20 fine-tuning epochs per round, and ptarget = 0.3.TabDDPM uses learning rate 10^-4 and batch size 32; the detector and diffusion-model settings are specified for the three experimental phases.
  • Evaluation protocol: The study evaluates Accuracy, Precision, Recall, F1, FPR, and ROC AUC, with P2 reusing the P1 test set and applying a fresh three-round schedule per method at τ = 0.5.Aggregates are standardized using StandardScaler fit on normal data only.

A. Evaluation Protocols for Reported F1

F1 performance is evaluated under five protocols, P1–P5, with Table III mapping each label to its corresponding evaluation setting.

  • A. Evaluation Protocols for Reported F1: F1 is reported under evaluation protocols P1–P5.
  • A. Evaluation Protocols for Reported F1: The protocols distinguish the evaluation settings used for reported F1 results.
  • A. Evaluation Protocols for Reported F1: Table III maps each protocol label to its evaluation setting.

VI. EXPERIMENTAL RESULTS AND DISCUSSION … D. ADT Recovery and Multiplier Retention

Across the Milano CDR evaluation, multiplier-based attacks produce near-perfect Phase 1 detection but realistic TabDDPM attacks expose severe robustness failures; ADT substantially restores detection while preserving strong performance on fixed-multiplier tests. The evaluation treats “5G” as a cellular-connectivity deployment context and defines realistic attacks as manifold-preserving rather than volumetric.

  • A. Evaluation Scope: The study evaluates the public Milano CDR corpus for reproducible comparison with Hussain et al., using “5G” as a deployment context rather than claiming standalone-5G-core trace collection.Coarse CDR aggregates are treated as generation-agnostic monitoring records, while no public labeled per-cell cellular DDoS CDR dataset exists.
  • B. Phase Progression Across Diff-DDoS: Phase 1 achieves 99.6%–100% F1 on multiplier-based attacks, but without retraining Phase 2 F1 collapses to near zero for SMS, Internet, and Blended and reaches 7.6% for ResNet50 silent-call.The Phase 1 result is expected because the handcrafted multipliers lie tens to hundreds of standard deviations from normal data.
  • C. Attack Realism of TabDDPM-Generated Attacks: Realistic attacks are defined as manifold-preserving samples within the statistical envelope of normal CDR traffic, targeting stealthy control-plane abuse rather than high-volume volumetric DDoS.Distributional fidelity is evaluated against the normal CDR aggregate manifold and compared with Phase 1 multiplier baselines.
  • C. Attack Realism of TabDDPM-Generated Attacks: Test attacks are assessed over 168 grids per scenario using target-channel mean absolute z-score as the primary realism metric and squared MMD2 with an RBF kernel as a distributional check.The comparison uses mean 3D aggregates from 40 attacked cells against 5,000 normal cell-level 3D features, with additional unconstrained TabDDPM draws.
  • D. ADT Recovery and Multiplier Retention: Architecture-specific recovery and multiplier-test retention are evaluated beyond Table IV using the TabDDPM protocol and protocol P5.The reported phase comparisons use τ = 0.5 unless otherwise stated.
  • D. ADT Recovery and Multiplier Retention: SimpleCNN ADT improves Internet F1 from 0% to 96.6% and Blended F1 to 79.6% at R3, but silent-call falls to 9.7% and SMS remains near 9%.Grad-CAM attributes the silent-call decline to suppressed attack scores on TabDDPM samples, with mean attacked-cell scores decreasing from 0.35 to 0.01 while localization is preserved.
  • D. ADT Recovery and Multiplier Retention: ADT converts inconsistent Blended per-cell accuracies into a near-uniform 9 × 9 grid at approximately 100% accuracy per cell.The heatmaps indicate that ADT resolves spatial inconsistencies across attacked locations.
  • D. ADT Recovery and Multiplier Retention: 100% F1 is retained by ResNet50 after ADT on SMS, silent-call, and Blended multiplier tests, while Internet F1 falls from 99.6% to 66.8% but rises to 86.3% at τ = 0.99.Internet attack-cell recall remains 100% (TP = 6720, FN = 0), so the decline reflects false positives on normal cells rather than missed multiplier attacks.

E. Comparison with Generative and Adversarial-Training Baselines · F. Inverse-Guidance Sensitivity and Online Latency

Across baseline comparisons, ADT provides the strongest or tied performance under both default and validation-calibrated thresholds, while uniquely combining manifold-preserving attack synthesis with robust blended detection. Its inverse-guidance realism is stable across tested settings, robustness is nonmonotonic, and online inference remains far below the 600-second inter-grid interval.

  • E. Comparison with Generative and Adversarial-Training Baselines: The baseline evaluation compares ADT with fixed-multiplier training, FGSM, PGD, and CTGAN using common three-round schedules on ResNet50 under distinct default and calibrated threshold protocols.Protocol P2 uses τ = 0.5, whereas P3 applies each method’s validation-selected τF1max once to the locked attack test.
  • E. Comparison with Generative and Adversarial-Training Baselines: At τ = 0.5, ADT reaches 95.1% Blended F1 versus ≤77.2% for FGSM, PGD, and CTGAN, leads Call at 80.4%, and ties Internet at 66.8%.These round-3 scores use the locked TabDDPM test and identical three-round iterative retraining schedules for the compared methods.
  • E. Comparison with Generative and Adversarial-Training Baselines: 100% SMS F1 is achieved by ADT, FGSM, and PGD after validation calibration, while CTGAN reaches 47.3%; Internet and Blended approach 100% for all methods.Call performance is essentially tied between ADT and PGD (90.5% vs. 89.5%, with overlapping bootstrap 95% confidence intervals).
  • E. Comparison with Generative and Adversarial-Training Baselines: ADT uniquely combines manifold-preserving standalone attack synthesis with deployable Blended robustness, whereas FGSM and PGD perturb existing grids and CTGAN lacks comparable calibrated SMS performance.CTGAN trains ≈21× faster than TabDDPM (62.5 s versus 1296.0 s) but shows Blended regression from 99.7% to 66.8% at τ = 0.5.
  • F. Inverse-Guidance Sensitivity and Online Latency: Inverse-guidance scale λ and target difficulty ptarget preserve target-channel MMD2 at approximately 0.51–0.57, but fresh-retrain robustness varies sensitively and nonmonotonically across settings.The sweep tests λ ∈ {0.25, 0.5, 1.0, 2.0} and ptarget ∈ {0.1, 0.3, 0.5, 0.7} on a locked TabDDPM attack test.
  • F. Inverse-Guidance Sensitivity and Online Latency: SimpleCNN averages 35.7 ms per grid with 49.5 ms p95, while ResNet50 averages 42.7 ms with 60.4 ms p95 on an RTX 4090.Each 9 × 9 × 3 grid aggregates a 10-minute window, making the operative CPS interval 600 seconds rather than per-packet or uRLLC-scale latency.
  • F. Inverse-Guidance Sensitivity and Online Latency: The reported latency target is periodic CDR-grid analytics: Diff-DDoS is not designed for per-packet or approximately 1 ms uRLLC-level inference.The latency measurements use batch 1, scenario averaging, 500 iterations, and 50 warmup iterations on an RTX 4090.

G. Operating Points: ROC-AUC and Threshold Calibration · H. Grad-CAM Interpretability of the SimpleCNN Silent-Call Anomaly · VII. LIMITATIONS AND CONCLUSION

The study shows that manifold-preserving attacks expose detector brittleness, while ADT with ResNet50 restores deployable robustness, but external validity remains limited by synthetic evaluation, coarse realism proxies, and Milano-era data. Operating-point analysis and Grad-CAM further reveal threshold sensitivity and SimpleCNN’s score–saliency failure under ADT.

  • G. Operating Points: ROC-AUC and Threshold Calibration: AUC reached 0.95–1.0 for ResNet50 after ADT, while SimpleCNN recovered to 0.85 on SMS and 0.96 on Blended from Phase 2 lows of 0.43 and 0.75.Phase 1 showed perfect separability with AUC=1.0, indicating that the AUC/F1 gap reflects the decision rule.
  • G. Operating Points: ROC-AUC and Threshold Calibration: Validation calibration selected τF1max and τY on 84 of 168 attack images and applied them once to 84 locked attack grids for SMS and silent-call.The locked evaluation was separate from the primary 335-image micro-averaged evaluation.
  • G. Operating Points: ROC-AUC and Threshold Calibration: Neither calibration nor cost-sensitive retraining raised SimpleCNN SMS F1 above 12.3 ± 8.6%, so ResNet50 was adopted as the deployable sparse-attack detector.For silent-call, weighted BCE raised mean locked-test AUC from 0.65 to 0.85 ± 0.09, but F1 at τF1max remained 48.5 ± 25.3%.
  • H. Grad-CAM Interpretability of the SimpleCNN Silent-Call Anomaly: SimpleCNN silent-call F1 fell from 52.6% in Phase 2 to 9.7% in Phase 3, despite Grad-CAM showing comparable attacked-cell saliency across phases.Across 40 TabDDPM silent-call grids, attacked-cell scores fell from 0.35 ± 0.47 to 0.01±0.07, while saliency mass remained 0.51±0.01 versus 0.51± 0.02.
  • H. Grad-CAM Interpretability of the SimpleCNN Silent-Call Anomaly: On multiplier-based attacks, mean scores stayed near 1.0 in both phases, but Phase 3 saliency became diffuse, declining from 0.51 ± 0.01 to 0.18 ± 0.24.This contrast supports the interpretation that ADT on SimpleCNN overfits the TabDDPM silent-call distribution rather than improving robust detection.
  • VII. LIMITATIONS AND CONCLUSION: 79.6% silent-call, 100% Internet, and 92.8% Blended F1 were recovered by ADT with ResNet50, while multiplier-trained detectors failed on manifold-preserving attacks.ADT retained near-perfect multiplier-test performance except an Internet calibration shift.
  • VII. LIMITATIONS AND CONCLUSION: External validity is constrained because no labeled per-cell cellular DDoS ground truth exists, realism uses coarse aggregate statistics, and Milano traffic reflects the 3G/4G era.These proxies omit full grid spatiotemporal structure and do not establish field adversary fidelity.
  • VII. LIMITATIONS AND CONCLUSION: Operator CDR validation and packet-level 5G corpora should test transfer beyond Milano, alongside future work on adaptive adversaries and regional scale-up.The paper presents pre-deployment stress testing as a practical safeguard for 5G-connected CPS and IoT services facing stealthy signaling abuse.
Loading 2608.17796v1…