Packet-Level Traffic Modeling with Heavy-Tailed Payload and Inter-Arrival Distributions for Digital Twins

Enes Koktas; Peter Rost

REVIEW 3 major objections 4 minor 20 references

Packet-Level Traffic Modeling with Heavy-Tailed Payload and Inter-Arrival Distributions for Digital Twins

T0 review · 3 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A compact generator combining a hidden Markov model and a Student-t mixture matches real packet traffic's size and timing on most traces, in about 0.2 MB.

desk verdict A compact HMM+Student-t MDN packet generator that is genuinely useful for RAN digital twins, with a clear likelihood and a fair evaluation design, but the headline comparative claim rests on single-run point estimates without uncertainty quantification. read the letter →

arxiv 2512.15432 v2 pith:B2IJNIJD submitted 2025-12-17 cs.NI

Enes Koktas , Peter Rost This is my paper · ORCID

classification cs.NI

keywords packet-leveltrafficgenerationdigitaltwinhiddenMarkovmodelmixturedensitynetworkStudent-tdistributionheavy-tailedinter-arrivaltimessimulationflowdiversity

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to show that a very small, interpretable generator can reproduce the packet-level statistics that matter for network digital twins: payload lengths and the gaps between packets, including their heavy tails. The proposed model chains a hidden Markov model—whose states capture regimes like buffering, streaming, and idle—onto a mixture density network that outputs a Student-t mixture for each state. The mixture's heavier tails and an explicit idle state anchored at the tail onset let the generator produce rare long gaps without inflating the whole model. Evaluated on web, smart home, and encrypted media traces, the generator matches marginal distributions, temporal autocorrelation, and flow diversity more closely than recent neural, transformer, and HMM baselines in most cases, while occupying about 0.2 MB and around 42,000 parameters. If this holds, digital twins could run realistic packet-level traffic without shipping raw traces or large models to the edge.

What carries the argument

The load-bearing mechanism is the coupling of a tail-aware HMM with a Student-t mixture density network. The HMM is initialized with k-means clusters plus an explicit idle state whose timing mean is placed at the empirical 99.8th percentile of the normalized log-IAT (υδ, equations 9–11); this anchors a basin for rare long gaps and prevents core-state variance inflation. The MDN then models p(z|state, flow-length) as a mixture of diagonal bivariate Student-t kernels (equation 32), with per-component degrees of freedom that let each component thicken its own tail. Together the state process drives dwell times and burstiness, while the Student-t mixture shapes the joint geometry of payload and

What would settle it

Train the generator on a trace that contains inter-arrival gaps beyond the one-hour cap (or whose log-space tail is distinctly non-Student-t, e.g., a power law with finite moments that Student-t cannot approximate), then check whether the synthetic IAT distribution reproduces those gaps; a failure to do so would confirm that the claimed heavy-tail fidelity is limited to the observed, capped range. Alternatively, run the model on a trace where the tail fraction above the 99.8th percentile is very high, so the idle-state anchor is misspecified, and compare CDFs.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that packet-level traffic can be generated accurately by factoring each flow into a small discrete state process and a per-state conditional density. The state process is a K-state HMM with diagonal Gaussian emissions, initialized so that one state—an 'idle' state—is anchored at a high quantile of the inter-arrival time distribution; this keeps core states from stretching their variances to cover rare long gaps. The emission model is a mixture density network that, given the one-hot state and the flow length, outputs a diagonal bivariate Student-t mixture (equation 32). The authors report that on four public traces this hybrid generator is the closest match

Load-bearing premise

The model assumes that the log-transformed, normalized space transforms real inter-arrival times and payloads into shapes that a diagonal Student-t mixture can represent, and that the tail anchor estimated from training flows (the 99.8th percentile of training IATs) adequately represents the tail beyond the training range; since preprocessing caps IATs at one hour and synthesis clips to observed ranges, the demonstrated heavy-tail fidelity only covers the capped, observed reg

Editorial extensions

If this is right

If the central claim holds, a network digital twin can carry a packet-level traffic model in a few hundred kilobytes and regenerate it locally as traffic changes, avoiding continuous transfer of raw traces.
The explicit idle state and Student-t kernels give the generator a handle on rare multi-second gaps, so simulated buffer drain, timeout, and queue occupancy may be more realistic than with Gaussian or regular-spacing generators.
Because the generator samples new state paths rather than replaying training states, it produces synthetic flows with variability closer to held-out test flows across the four evaluated datasets.
The separation of the HMM state process from the MDN emission layer means recalibration reduces to re-estimating a small transition matrix and re-fitting a compact feed-forward network, which fits the digital-twin loop's need for low-overhead adaptation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The two-stage design—regime HMM plus heavy-tailed emission—is not specific to packets; the same machinery could model other heavy-tailed time series with regime switches, such as web requests, transaction arrivals, or failure logs, wherever an explicit idle state maps to a natural tail anchor.
The paper clips synthesis to the operational ranges used in preprocessing; a testable extension would be to let the idle-state degrees of freedom extrapolate beyond the training cap and measure whether long-gap statistics (e.g., gaps beyond one hour) are preserved, which the current evaluation does not cover.
Because the MDN conditions on flow length, the generator implicitly learns session-scale structure; this suggests that incorporating other cheap per-flow covariates (e.g., flow direction or port class) could improve cross-flow diversity at negligible memory cost.
The reported runtime for the Student-t likelihood is higher than the transformer baseline on the largest trace; an optimization pass (e.g., precomputing gamma terms) would strengthen the deployment story for latency-sensitive digital-twin loops.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

This is a solid incremental engineering paper, not a breakthrough. What it does well is concrete: it replaces the Gaussian MDN kernels from the authors' earlier work with Student-t kernels, anchors an explicit idle HMM state at a tail quantile, and conditions on flow length. The result is a roughly 0.16 MB generator that fits marginal CDFs, autocorrelation, and flow-level Wasserstein distances better than larger baselines in most of Table II's cells. The model is cleanly specified: the likelihood in Eq. (32) is legitimate, the HMM initialization is sensible, and the training/evaluation split is flow-wise, which makes the per-flow metrics paired and appropriate. Using public traces is a plus, and the paper is honest about what is tunable and what is not.

The soft spots are real, but they are mostly about the strength of the evidence, not the architecture. Table II reports one run per model/dataset/metric. There are no error bars, no multiple seeds, no significance tests. Several of the advantages are small—HTTP AC IAT 0.04 versus 0.05–0.06, and a tie at 0.02 on WD IAT. Combined with the fact that the baselines are reimplemented "to the best of our knowledge" and the transformer needs temperature 1.5 to look good, the central phrase "matches more closely in most cases" is a point estimate, not a statistically supported ranking. That is the load-bearing weakness, and it is fixable.

The tail-cap concern is secondary. The authors cap IATs at one hour and clip synthesis to observed ranges, so they do not demonstrate extrapolation beyond the training tail. But the model is explicitly designed for per-deployment recalibration, and the evaluation is inside the observed range, so this is a limitation of scope rather than a flaw in the comparative claim. There is also no circularity problem: the tail anchor comes from training flows, and evaluation is on held-out flows.

Who is this for? Practitioners building packet-level RAN digital twins who want a compact, interpretable generator they can retrain locally. If that is your context, this is a plausible default to try, and the math is clean enough to reimplement without code.

My recommendation: send it to peer review. It deserves referee time. The main ask should be uncertainty quantification—multiple seeds or bootstrap confidence intervals on Table II—and ideally code release. With that, it becomes a dependable accepted paper rather than a promising preprint.

Referee Report

3 major / 4 minor

Summary. The paper proposes a compact packet-level traffic generator for network digital twins, combining a small hidden Markov model (HMM) with a mixture density network (MDN) that emits diagonal bivariate Student-t mixtures for payload length and inter-arrival time (IAT). The HMM is initialized with a tail-anchored idle state and Dirichlet priors, and the MDN is trained on soft state posteriors. The model is evaluated on four public traces (HTTP, UDP, Facebook audio, Facebook video) against several neural, transformer, and HMM baselines, reporting average per-flow CDFs, autocorrelation errors, and Wasserstein distances in Table II. The paper claims that the proposed generator matches real traffic most closely in most cases while using orders of magnitude fewer parameters (around 0.2 MB).

Significance. If the comparative claims are supported, the work is a useful contribution to network digital twins: it provides a compact, interpretable, and potentially easily recalibrated generator that explicitly targets heavy-tailed IATs and payloads, which is important for buffer and delay dynamics. The design choices (idle state anchored at a high empirical quantile, Student-t kernels, flow-length conditioning) are well motivated, and the evaluation on public traces with a flow-wise train/test split is a strength. The explicit reporting of parameter counts and memory footprint is also useful. However, the headline comparative claim rests on Table II, which currently reports single-run point estimates without uncertainty quantification; this is the main weakness.

major comments (3)

[Section IV / Table II] The central claim that the proposed generator 'matches ... more closely ... in most cases' is supported only by single-run point estimates in Table II. There are no standard errors, confidence intervals, or significance tests, and each cell corresponds to one train/test split. Several advantages are small (e.g., HTTP AC IAT 0.04 vs 0.05–0.06 for [10]/[17]; HTTP WD IAT 0.02 ties with [10]/[17]). Under another split or random seed these cells could easily flip. Please provide repeated evaluations (multiple seeds or flow-bootstrap) with confidence intervals, or a paired statistical test over flows, to substantiate the comparative claim.
[Section IV, baseline implementations and [9] temperature] The baselines are described as implemented 'to the best of our knowledge' (Section IV), which introduces uncontrolled implementation uncertainty. In particular, the transformer [9] is evaluated at temperature 1.0 and at 1.5, with 1.5 used in the final comparison and reported as giving better marginal fits. This hand-tuned temperature could affect the ranking; a sensitivity analysis over temperatures, or use of the official implementation/outputs, is needed to make the comparison fair and reproducible.
[Sections II and III-C] The preprocessing caps IATs at one hour and payloads at the Ethernet MTU (Section II), and synthesis clips samples to 'observed operational ranges' (Section III-C). The tail anchor tau_delta is estimated from the empirical 99.8th percentile of training flows (Eq. 9). Consequently, the claimed heavy-tail fidelity is demonstrated only inside the observed, capped range; the model cannot generate gaps beyond the training cap and extrapolation to untested tails is not evaluated. Please state this limitation explicitly and, ideally, assess sensitivity to the cap and threshold choices.

minor comments (4)

[Section III-B] The phrase 'discussed in Section III–III-B' appears to be a formatting error; should likely read 'Section III-B'.
[Abstract vs Section IV] The abstract says the model occupies 'around 0.2 MB', while Section IV reports 'around 0.164 MB' and later '0.164 MB'. Please unify the numbers.
[Table II] The table would be easier to read with consistent decimal places and, ideally, error bars or superscripts indicating statistical significance (if added).
[Section IV / Figure 5] The zoomed IAT CDF in Fig. 5 begins at y=0.8, which may exaggerate differences; consider showing the full range in the main plot or a separate inset.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: model fit on training flows and evaluated on held-out flows, with self-citations only as motivation.

full rationale

The paper's derivation chain is self-contained: the HMM/MDN model is trained on training-flow packets (normalization moments in Eq. 4, tail threshold in Eq. 9, HMM posteriors, and MDN likelihood in Eq. 34), and all reported CDF, AC, and Wasserstein metrics are computed on held-out test flows whose raw packets were not used for fitting. No predicted quantity is fed back as an input. The flow-length conditioning is motivated by the authors' previous figure in [15], but this is a design choice, not a result derived from that citation; the current paper independently states the observed short-flow difficulty and uses held-out evaluation. The tail anchor and idle-state activation are estimated from training flows only, so they do not encode test-flow outcomes. Baselines are reimplemented by the authors and the comparison lacks error bars, but that is an evidentiary limitation affecting the strength of the comparative claim, not circularity. No equation reduces to its own input, and no fitted parameter is renamed as a prediction. Therefore the paper exhibits no significant circularity.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

No new physical entities, forces, or conserved quantities are introduced. The model's latent states are conventional HMM constructs. The load-bearing free choices are the state count Kcore, the tail activation threshold, the tail-anchor quantile, the transition priors, and the MDN size/floors; none are derived from first principles.

free parameters (5)

Kcore (number of core HMM states) = typically 2–4, set by practitioner or 'light model-selection rule'
The state cardinality is not learned; it is a tunable hyperparameter that controls how many behavioral states the HMM can represent (Section III-A).
theta_tail (idle activation threshold) = not reported
The idle state is activated if the fraction of packets above upsilon_delta exceeds this threshold (Section III-A); the value is a hand-chosen hyperparameter.
tau_delta quantile (99.8th percentile of log IAT) = empirical 99.8th percentile on training flows
The idle-state mean is anchored at this data-derived value (Eqs. 9–11); it controls where the tail state sits and is not derived from a principle.
Dirichlet pseudo-counts lambda_self, lambda_off, lambda_idle, lambda_leak = not reported numerically
These transition-prior hyperparameters stabilize EM and shape dwell times, especially for the idle state (Eq. 24). The paper states the same prior is used across datasets but gives no values.
MDN size (M=32 components, H=128 hidden units) and variance/DOF floors = M=32, H=128, epsilon_0>0, nu_m>1
The network width, mixture count, and numerical floors are fixed across datasets by hand (Sections III-B, III-C); their values affect the fidelity/bias trade-off.

assumptions (5)

domain assumption Packet traffic within a flow is generated by a first-order Markov chain over a small set of latent states (Eq. 12).
The core modeling premise: buffering, streaming, and idle behaviors can be represented as a finite-state Markov process with geometric dwell times (Eq. 25).
domain assumption After log transform and moment normalization, the joint distribution of payload and IAT in each state is well approximated by a diagonal bivariate Student-t mixture (Eqs. 31–32).
The Student-t family and the diagonal covariance form are chosen for tractability, not derived from trace statistics; the same nu_m is shared across coordinates, limiting tail-shape flexibility.
ad hoc to paper Anchoring the idle state at the empirical 99.8th percentile of log-IAT separates the heavy tail from core states without discarding extremes (Eqs. 9–11).
This initialization is a paper-specific heuristic; it is motivated by observed failure modes but is not justified by a formal tail model.
domain assumption Capping IATs at one hour and payloads at the Ethernet MTU preserves the statistics needed for DT evaluation.
Section II caps values during preprocessing and Section III-C clips generated samples to the same ranges, so any traffic beyond these bounds is not represented in the model or the metrics.
standard math The Student-t density formula and normalizing constant (Eqs. 28–29) from Lange et al. [16] are correct.
The likelihood derivation relies on the standard univariate Student-t density; this is a mathematical background result.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Packet-Level Traffic Modeling with Heavy-Tailed Payload and Inter-Arrival Distributions for Digital Twins." pith.science (2026). https://pith.science/paper/B2IJNIJD

@misc{pith2026251215432,
  author       = {Pith},
  title        = {Pith review of: Packet-Level Traffic Modeling with Heavy-Tailed Payload and Inter-Arrival Distributions for Digital Twins},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B2IJNIJD}},
  note         = {Machine review of arXiv:2512.15432}
}

read the original abstract

Digital twins of radio access networks require packet-level traffic generators that reproduce the size and timing of packets while remaining compact and easy to recalibrate as traffic changes. We address this need with a hybrid generator that combines a small hidden Markov model, which captures buffering, streaming, and idle states, with a mixture density network that models the joint distribution of payload length and inter-arrival time (IAT) in each state using Student-t mixtures. The state space and emission family are designed to handle heavy-tailed IAT by anchoring an explicit idle state in the tail and allowing each component to adapt its tail thickness. We evaluate the model on public traces of web, smart home, and encrypted media traffic and compare it with recent neural network and transformer based generators as well as hidden Markov baselines. Across most datasets and metrics, including average per-flow cumulative distribution functions, autocorrelation based measures of temporal structure, and Wasserstein distances between flow descriptors, the proposed generator matches the real traffic most closely in the majority of cases while using orders of magnitude fewer parameters. The full model occupies around 0.2 MB in our experiments, which makes it suitable for deployment inside digital twins where memory footprint and low-overhead adaptation are critical.

Figures

Figures reproduced from arXiv: 2512.15432 by the authors.

**Figure 2.** Average per-flow CDF comparison of payload length [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 4.** Average per-flow CDF comparison of payload length [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

Figures from the paper (1 more)

**Figure 5.** Figure 5: Average per-flow CDF comparison of payload length [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 1 canonical work pages

[10]

IP traffic generator based on hidden Markov models,

H. Red ˇzovi´c, A. Smiljani´c, and M. Bjelica, “IP traffic generator based on hidden Markov models,” inProc. of 4th Int. Conf. on Electr., Electron. and Comput. Eng. (IcETRAN), Jun 2017, pp. TEI2.3.1–6

2017
[17]

Time-series generative adversarial networks,

J. Yoon, D. Jarrett, and M. van der Schaar, “Time-series generative adversarial networks,” inAdv. in Neural Inf. Process. Syst., H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alch ´e-Buc, E. Fox, and R. Garnett, Eds., vol. 32. Curran Associates, Inc.,
[9]

TrafficGPT: Breaking the token barrier for efficient long traffic analysis and generation,

J. Qu, X. Ma, and J. Li, “TrafficGPT: Breaking the token barrier for efficient long traffic analysis and generation,” 2024. [Online]. Available: https://arxiv.org/abs/2403.05822

arXiv 2024
[1]

Digital twins: A survey on enabling technologies, challenges, trends and future prospects,

S. Mihai, M. Yaqoob, D. V . Hung, W. Davis, P. Towakel, M. Raza, M. Karamanoglu, B. Barn, D. Shetve, R. V . Prasad, H. Venkataraman, R. Trestian, and H. X. Nguyen, “Digital twins: A survey on enabling technologies, challenges, trends and future prospects,”IEEE Commun. Surveys Tuts., vol. 24, no. 4, pp. 2255–2291, 2022

2022
[2]

Network digital twin: Concepts and reference architecture,

C. Zhou, H. Yang, X. Duan, D. Lopez, A. Pastor, Q. Wu, M. Boucadair, and C. Jacquenet, “Network digital twin: Concepts and reference architecture,” Internet Engineering Task Force, Tech. Rep., Mar. 2024, work in Progress. [Online]. Available: https: //datatracker.ietf.org/doc/draft-irtf-nmrg-network-digital-twin-arch/05/

2024
[3]

Self-similarity in world wide web traffic: evidence and possible causes,

M. Crovella and A. Bestavros, “Self-similarity in world wide web traffic: evidence and possible causes,”IEEE/ACM Trans. Netw., vol. 5, no. 6, pp. 835–846, 1997

1997
[4]

Feasibility of state space models for network traffic generation,

A. Chu, X. Jiang, S. Liu, A. Bhagoji, F. Bronzino, P. Schmitt, and N. Feamster, “Feasibility of state space models for network traffic generation,” inProc. of the 2024 SIGCOMM Workshop on Networks for AI Comput., ser. NAIC ’24. New York, NY , USA: Association for Computing Machinery, 2024, p. 9–17. [Online]. Available: https://doi.org/10.1145/3672198.3673792

arXiv 2024
[5]

Using GANs for sharing networked time series data: Challenges, initial promise, and open questions,

Z. Lin, A. Jain, C. Wang, G. Fanti, and V . Sekar, “Using GANs for sharing networked time series data: Challenges, initial promise, and open questions,” inProc. of the ACM Internet Meas. Conf., ser. IMC ’20. New York, NY , USA: Association for Computing Machinery, 2020, p. 464–483. [Online]. Available: https://doi.org/10.1145/3419394.3423643

arXiv 2020

Show all 20 references

[6]

Mobile user traffic generation via multi-scale hierarchical GAN,

T. Li, S. Hui, S. Zhang, H. Wang, Y . Zhang, P. Hui, D. Jin, and Y . Li, “Mobile user traffic generation via multi-scale hierarchical GAN,” ACM Trans. Knowl. Discov. Data, vol. 18, no. 8, Jul. 2024. [Online]. Available: https://doi.org/10.1145/3664655

2024 doi
[7]

Generative deep learning for internet of things network traffic generation,

M. R. Shahid, G. Blanc, H. Jmila, Z. Zhang, and H. Debar, “Generative deep learning for internet of things network traffic generation,” in2020 IEEE 25th Pacific Rim Int. Symp. on Dependable Comput. (PRDC), 2020, pp. 70–79

2020
[8]

NeCSTGen: An ap- proach for realistic network traffic generation using deep learning,

F. Meslet-Millet, S. Mouysset, and E. Chaput, “NeCSTGen: An ap- proach for realistic network traffic generation using deep learning,” in IEEE Global Commun. Conf. (GLOBECOM), 2022, pp. 3108–3113

2022
[11]

Characterization of encrypted and VPN traffic using time-related,

G. Draper-Gil, A. H. Lashkari, M. S. I. Mamun, and A. A. Ghorbani, “Characterization of encrypted and VPN traffic using time-related,” in Proc. of the 2nd Int. Conf. on Inf. Syst. Secur. and Privacy (ICISSP), 2016, pp. 407–414

2016
[12]

A tutorial on hidden markov models and selected applica- tions in speech recognition,

L. Rabiner, “A tutorial on hidden markov models and selected applica- tions in speech recognition,”Proc. IEEE, vol. 77, no. 2, pp. 257–286, 1989

1989
[13]

hmmlearn: Hidden markov models in python,

D. Cournapeau, F. Pedregosa, G. Varoquaux, S. Lebedev, A. Lee, and M. Danielson, “hmmlearn: Hidden markov models in python,” 2015, accessed Nov. 12, 2025. [Online]. Available: https://github.com/ hmmlearn/hmmlearn 10

2015
[14]

On the self-similar nature of ethernet traffic (extended version),

W. Leland, M. Taqqu, W. Willinger, and D. Wilson, “On the self-similar nature of ethernet traffic (extended version),”IEEE/ACM Trans. Netw., vol. 2, no. 1, pp. 1–15, 1994

1994
[15]

State aware traffic generation for real-time network digital twins,

E. Koktas and P. Rost, “State aware traffic generation for real-time network digital twins,” in2025 IEEE 36th IEEE Int. Symp. Pers. Indoor Mob. Radio Commun. (PIMRC), 2025, pp. 1–6

2025
[16]

Robust statistical modeling using the t distribution,

K. L. Lange, R. J. A. Little, and J. M. G. Taylor, “Robust statistical modeling using the t distribution,”J. Amer. Statist. Assoc., vol. 84, no. 408, pp. 881–896, 1989. [Online]. Available: https://doi.org/10.1080/01621459.1989.10478852

1989
[18]

Wide area traffic: the failure of poisson modeling,

V . Paxson and S. Floyd, “Wide area traffic: the failure of poisson modeling,”IEEE/ACM Trans. on Netw., vol. 3, no. 3, pp. 226–244, 1995

1995
[19]

Computational optimal transport,

G. Peyr ´e and M. Cuturi, “Computational optimal transport,” 2020. [Online]. Available: https://arxiv.org/abs/1803.00567

2020 arXiv
[2019]

Available: https://proceedings.neurips.cc/paper files/ paper/2019/file/c9efe5f26cd17ba6216bbe2a7d26d490-Paper.pdf

[Online]. Available: https://proceedings.neurips.cc/paper files/ paper/2019/file/c9efe5f26cd17ba6216bbe2a7d26d490-Paper.pdf

2019

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [10]

IP traffic generator based on hidden Markov models,

H. Red ˇzovi´c, A. Smiljani´c, and M. Bjelica, “IP traffic generator based on hidden Markov models,” inProc. of 4th Int. Conf. on Electr., Electron. and Comput. Eng. (IcETRAN), Jun 2017, pp. TEI2.3.1–6

2017

[2] [17]

Time-series generative adversarial networks,

J. Yoon, D. Jarrett, and M. van der Schaar, “Time-series generative adversarial networks,” inAdv. in Neural Inf. Process. Syst., H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alch ´e-Buc, E. Fox, and R. Garnett, Eds., vol. 32. Curran Associates, Inc.,

[3] [9]

TrafficGPT: Breaking the token barrier for efficient long traffic analysis and generation,

J. Qu, X. Ma, and J. Li, “TrafficGPT: Breaking the token barrier for efficient long traffic analysis and generation,” 2024. [Online]. Available: https://arxiv.org/abs/2403.05822

arXiv 2024

[4] [1]

Digital twins: A survey on enabling technologies, challenges, trends and future prospects,

S. Mihai, M. Yaqoob, D. V . Hung, W. Davis, P. Towakel, M. Raza, M. Karamanoglu, B. Barn, D. Shetve, R. V . Prasad, H. Venkataraman, R. Trestian, and H. X. Nguyen, “Digital twins: A survey on enabling technologies, challenges, trends and future prospects,”IEEE Commun. Surveys Tuts., vol. 24, no. 4, pp. 2255–2291, 2022

2022

[5] [2]

Network digital twin: Concepts and reference architecture,

C. Zhou, H. Yang, X. Duan, D. Lopez, A. Pastor, Q. Wu, M. Boucadair, and C. Jacquenet, “Network digital twin: Concepts and reference architecture,” Internet Engineering Task Force, Tech. Rep., Mar. 2024, work in Progress. [Online]. Available: https: //datatracker.ietf.org/doc/draft-irtf-nmrg-network-digital-twin-arch/05/

2024

[6] [3]

Self-similarity in world wide web traffic: evidence and possible causes,

M. Crovella and A. Bestavros, “Self-similarity in world wide web traffic: evidence and possible causes,”IEEE/ACM Trans. Netw., vol. 5, no. 6, pp. 835–846, 1997

1997

[7] [4]

Feasibility of state space models for network traffic generation,

A. Chu, X. Jiang, S. Liu, A. Bhagoji, F. Bronzino, P. Schmitt, and N. Feamster, “Feasibility of state space models for network traffic generation,” inProc. of the 2024 SIGCOMM Workshop on Networks for AI Comput., ser. NAIC ’24. New York, NY , USA: Association for Computing Machinery, 2024, p. 9–17. [Online]. Available: https://doi.org/10.1145/3672198.3673792

arXiv 2024

[8] [5]

Using GANs for sharing networked time series data: Challenges, initial promise, and open questions,

Z. Lin, A. Jain, C. Wang, G. Fanti, and V . Sekar, “Using GANs for sharing networked time series data: Challenges, initial promise, and open questions,” inProc. of the ACM Internet Meas. Conf., ser. IMC ’20. New York, NY , USA: Association for Computing Machinery, 2020, p. 464–483. [Online]. Available: https://doi.org/10.1145/3419394.3423643

arXiv 2020