TraGe: A Generic Packet Representation for Traffic Classification Based on Header-Payload Differences

Chungang Lin; Tianyu Zuo; Weiyao Zhang; Xuying Meng; Yilong Jiang; Yujun Zhang

REVIEW 2 major objections 1 minor 25 references

TraGe: A Generic Packet Representation for Traffic Classification Based on Header-Payload Differences

T0 review · 2 major / 1 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read TraGe learns a generic packet representation by pre-training header and payload with different masking tasks, and reports F1 scores of 0.7484 and 0.9331 on two ISCX-VPN classification tasks.

desk verdict TraGe's header/payload differentiated masking is a solid idea with plausible results, but the paper never defines where the header ends and the payload begins, which is the load-bearing detail that needs fixing. read the letter →

arxiv 2506.14151 v2 pith:LMJIZHL3 submitted 2025-06-17 cs.NI

Chungang Lin , Yilong Jiang , Weiyao Zhang , Xuying Meng , Tianyu Zuo , Yujun Zhang This is my paper · ORCID

classification cs.NI

keywords trafficclassificationgenericpacketrepresentationpre-trainingheader-payloaddifferentiationfield-levelmaskingdynamicencryptedTransformer

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 proposes TraGe, a pre-trained model that learns a generic packet representation by treating a network packet's header and payload differently during pre-training. The core claim is that because header bytes are continuously structured protocol fields while payload bytes are effectively discontinuous, two separate masked-language-modeling tasks—field-level masking for the header and random masking for the payload—yield a representation that transfers to traffic classification with limited labeled data. On the ISCX-VPN benchmark, TraGe reports F1 of 0.7484 across 17 application classes and 0.9331 across 12 service classes, outperforming all twelve baselines with up to a 6.97% gain over the best pre-trained model. The paper also reports ablations showing that both field-level masking and dynamic masking contribute to the gains.

What carries the argument

The mechanism is header-payload differentiated pre-training: two masked-language-modeling tasks share one Transformer encoder. For header bytes, MLM-FM masks contiguous runs whose lengths are drawn from a geometric distribution $\ell \sim \mathrm{Geo}(p)$ (with $p=0.7$), approximating whole protocol fields without parsing; for payload bytes, MLM-RM masks randomly scattered tokens because payload bytes are treated as discontinuous. Dynamic masking regenerates the masked positions during training so the model does not overfit to fixed byte positions.

What would settle it

Measure header-field lengths across a broad packet corpus and test whether their distribution matches $\mathrm{Geo}(p)$ closely enough to justify the approximation; alternatively, compare TraGe's F1 with a variant that masks exactly parsed header fields. If geometric masking does not beat random masking on held-out data, or if parsed-field masking clearly outperforms geometric masking, the paper's central mechanism is falsified.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the byte-distribution difference between header and payload is a feature to exploit, rather than an obstacle. Existing pre-trained models borrow masked-language or masked-image tasks from text and vision, masking header and payload uniformly, which breaks the contiguous protocol fields that carry traffic meaning. TraGe instead assigns each packet half its own masking task: header bytes are masked in contiguous runs whose lengths follow a geometric distribution, approximating whole protocol fields without explicit parsing; payload bytes are masked at random, reflecting their fragmented or encrypted structure. Combined with dynamic masking that re-samples masked positions during training, this produces a generic packet representation that a small fine-tuning head can use for both application classification and service identification.

Load-bearing premise

The load-bearing premise is that a contiguous run of bytes whose length is drawn from a geometric distribution approximates a real protocol header field well enough that masking such runs teaches the model header structure; if real header fields are not geometrically distributed, or if the runs routinely cut across field boundaries, the field-level masking benefit asserted in the ablation is not established.

Editorial extensions

If this is right

A single pre-trained packet encoder can serve multiple downstream traffic classification tasks via light fine-tuning, reducing reliance on large labeled datasets.
Header-aware field-level masking preserves protocol byte continuity, so gains are achievable on encrypted and VPN traffic without deep packet parsing.
Dynamic masking prevents overfitting to fixed byte positions, making the learned representation more robust across different sampling configurations.
Because performance is stable across geometric-distribution parameter values, practitioners can adopt the method without per-dataset tuning of the masking distribution.

Reading between the lines

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

The geometric run-length assumption could be tested directly by comparing TraGe against a variant that masks precisely parsed header fields; if parsed-field masking clearly outperforms geometric masking, field alignment rather than run continuity is the driving factor.
The header/payload split may transfer to other semi-structured byte streams, such as encrypted DNS or QUIC flows, where header fields are known and stable.
Dynamic masking could be a general benefit for any byte-sequence encoder, and the paper's robustness results hint that mask scheduling matters more than the exact mask distribution.
Since the paper reports classification only on ISCX-VPN, applying the same pre-trained encoder to the other collected datasets (USTC-TFC and CIC-IoT) would test whether the generic representation truly generalizes beyond VPN traffic.

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

TraGe is a pre-training method for packet representations. The new thing is that it handles header and payload bytes differently: field-level masking on the header with lengths sampled from a geometric distribution, random masking on the payload, plus dynamic masking. That combination is not in ET-BERT, PERT, NetGPT, or YaTC. For that reason alone it deserves a look.

What the paper does well: the geometric-distribution masking is a pragmatic solution to the cost of parsing headers. The ablation in Table IV shows that removing field-level masking costs 5.88 F1 points on app classification, which suggests the effect is real, not a rounding error. The parameter sensitivity and sampling variability analyses are also useful, though they only cover TraGe itself, not the baselines.

The soft spots are real. The biggest is that the paper never says how the input byte sequence is divided into a header region and a payload region. Fig. 1 shows separate sequences, and the text describes two pre-training tasks, but no boundary rule is given. Header lengths depend on protocol fields, IHL, TCP data offset, and options. Without this, the input to MLM-FM is under-specified, and the attribution of the ablation gains to 'header continuity' is not established. This is more fundamental than the geometric approximation; even if geometric masking were perfect, we don't know which bytes it masks.

Other issues: no error bars or variance on the headline results (Tables II and III), and the split of the datasets into pre-training and fine-tuning is not described. The reader's concern about same-distribution leakage is real. No code or data release either. These are addressable, but they are current limits on verification.

I do not see any circularity problem. The claim is empirical, and the baselines are standard. NetGPT as a baseline is fine since two authors wrote it; the central claim doesn't depend on beating it.

Who is this for: people building traffic-classification pipelines or self-supervised representations for packets. It is not a theoretical breakthrough but an incremental, useful engineering idea. With the boundary defined and a few robustness fixes, it would be a decent conference paper. As is, I would not rely on the specific numbers until the split is clarified. I'd send it to peer review with a request for revision rather than desk-reject; the idea is sound and the evaluation is mostly honest.

Referee Report

2 major / 1 minor

Summary. TraGe is a pre-trained packet representation model for traffic classification. The paper proposes header-payload differentiated pre-training: a field-level masked language modeling task (MLM-FM) over the packet header, a random-masking task (MLM-RM) over the payload, and a dynamic masking strategy. After pre-training on portions of ISCX-VPN, USTC-TFC, and CIC-IoT, the model is fine-tuned on 17-class application classification and 12-class service identification on ISCX-VPN. The authors report that TraGe outperforms twelve baselines (statistical, deep-learning, and pre-trained), with F1-scores of 0.7484 and 0.9331 on the two tasks, and they include an ablation study, a parameter-sensitivity analysis, and a sampling-variability analysis.

Significance. If the results hold, TraGe would be a useful contribution: it is one of the few pre-training methods that explicitly differentiates header and payload structure rather than treating packet bytes as a homogeneous token sequence. The paper's strengths are its broad baseline comparison, the inclusion of an ablation isolating field-level masking and dynamic masking, and the robustness checks across masking parameters and sampling seeds. These are valuable in a field where many proposed representations are evaluated only once. The main limitation is that the core mechanism is not fully specified: the header/payload boundary is never defined, and the pre-training/fine-tuning data split is not described. Without those details, the reported gains cannot be independently verified, even though the overall idea is plausible and the experimental framework is mostly sound.

major comments (2)

[III-A.a, Fig. 1] The paper never defines how packet bytes are partitioned into header and payload regions. MLM-FM (Eq. 1) operates on the header and MLM-RM (Eq. 2) on the payload, and Fig. 1 shows separate byte sequences for the two regions, but the text does not state where the header ends and the payload begins. Header lengths are protocol-dependent (e.g., IPv4 IHL, TCP data offset, Ethernet header size, and VPN encapsulation in ISCX-VPN), so a fixed byte offset cannot be assumed without stating it. The paper explicitly avoids parsing for the masking step ('parsing packets can be computationally expensive', Section III-A.a), yet no alternative rule is given for the initial header/payload split. Consequently, the input to MLM-FM is undefined, and the ablation results in Table IV cannot be attributed to preserving header-field continuity. Please specify the exact boundary rule used, explain how variable-length header fields are handled, and state whether the rule is applied consistently across all three datasets.
[IV-A.a] The paper never defines how packet bytes are partitioned into header and payload regions. MLM-FM (Eq. 1) operates on the header and MLM-RM (Eq. 2) on the payload, and Fig. 1 shows separate byte sequences for the two regions, but the text does not state where the header ends and the payload begins. Header lengths are protocol-dependent (e.g., IPv4 IHL, TCP data offset, Ethernet header size, and VPN encapsulation in ISCX-VPN), so a fixed byte offset cannot be assumed without stating it. The paper explicitly avoids parsing for the masking step ('parsing packets can be computationally expensive', Section III-A.a), yet no alternative rule is given for the initial header/payload split. Consequently, the input to MLM-FM is undefined, and the ablation results in Table IV cannot be attributed to preserving header-field continuity. Please specify the exact boundary rule used, explain how variable-length header fields are handled, and state whether the rule is applied consistently across all three datasets.

minor comments (1)

[Abstract and Section I] Section IV-C says removing field-level masking causes a 5.88% decrease in F1-score; this is computed from Table IV as (0.7484-0.7044)/0.7484, but the corresponding decreases for Precision and Recall differ, so the text should state explicitly which metric the percentage refers to.

Circularity Check

0 steps flagged · score 2.0 of 10

No circularity: the central results are empirical comparisons against external baselines; the only self-citation (NetGPT as a baseline) is not load-bearing.

full rationale

TraGe is an empirical systems paper: the claimed F1 scores are measured outcomes on public datasets (ISCX-VPN), not quantities derived from the model's own definitions. The central comparison is against externally developed baselines (PERT, ET-BERT, YaTC, FlowPrint, etc.), so the core claim does not reduce to the paper's inputs. The only self-citation is NetGPT [17], authored by two of the TraGe authors (X. Meng and Y. Zhang), and it appears solely as one of four pre-trained baselines in Tables II and III; the claimed improvement over NetGPT is an empirical comparison, not a premise whose validity rests on the cited paper. The geometric-distribution masking is an approximation justified by a case study on header field lengths, not a fitted parameter renamed as a prediction, because the downstream classification labels are not defined in terms of the masking lengths. The paper does leave the header/payload boundary unspecified, but that is a reproducibility and rigor concern rather than circularity: no equation in the paper is equal to an input by construction, and no result is forced by a self-citation chain.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claim depends on the geometric-distribution approximation for header field lengths, which involves a free parameter p and a domain assumption, and on the continuity versus discontinuity dichotomy between header and payload. No new physical or conceptual entity is introduced.

free parameters (1)

geometric distribution parameter p = 0.7
Chosen by hand in Section IV-A.b to shape field-level masking lengths. The paper claims robustness across values, but the specific value is set manually and affects the masking behavior.

assumptions (3)

domain assumption Protocol header field lengths in real traffic follow a geometric distribution, so sampling a contiguous run length from Geo(p) approximates masking entire protocol fields.
Invoked in Section III-A.a and supported only by the ISCX-VPN case study in Fig. 2. The approximation ignores field start positions and other protocols.
domain assumption Payload bytes have no meaningful local continuity because of fragmentation and encryption.
Invoked in Section III-A.b to justify random masking for the payload.
domain assumption BERT-style transformer with 2-byte tokenization is a suitable encoder for raw packet bytes.
Adopted from ET-BERT in Section III-A.a and Section IV-A.b without independent justification.

how reviews work

0 comments

Cite this review

Pith. "Pith review of TraGe: A Generic Packet Representation for Traffic Classification Based on Header-Payload Differences." pith.science (2026). https://pith.science/paper/LMJIZHL3

@misc{pith2026250614151,
  author       = {Pith},
  title        = {Pith review of: TraGe: A Generic Packet Representation for Traffic Classification Based on Header-Payload Differences},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LMJIZHL3}},
  note         = {Machine review of arXiv:2506.14151}
}

read the original abstract

Traffic classification has a significant impact on maintaining the Quality of Service (QoS) of the network. Since traditional methods heavily rely on feature extraction and large scale labeled data, some recent pre-trained models manage to reduce the dependency by utilizing different pre-training tasks to train generic representations for network packets. However, existing pre-trained models typically adopt pre-training tasks developed for image or text data, which are not tailored to traffic data. As a result, the obtained traffic representations fail to fully reflect the information contained in the traffic, and may even disrupt the protocol information. To address this, we propose TraGe, a novel generic packet representation model for traffic classification. Based on the differences between the header and payload-the two fundamental components of a network packet-we perform differentiated pre-training according to the byte sequence variations (continuous in the header vs. discontinuous in the payload). A dynamic masking strategy is further introduced to prevent overfitting to fixed byte positions. Once the generic packet representation is obtained, TraGe can be finetuned for diverse traffic classification tasks using limited labeled data. Experimental results demonstrate that TraGe significantly outperforms state-of-the-art methods on two traffic classification tasks, with up to a 6.97% performance improvement. Moreover, TraGe exhibits superior robustness under parameter fluctuations and variations in sampling configurations.

Figures

Figures reproduced from arXiv: 2506.14151 by the authors.

**Figure 1.** The framework of TraGe. packet representation. However, the random masking approach disrupts the continuity of the header fields. For example, a byte sequence like “b11eac20” may represent specific network information, such as a sequence number. Random masking could mask only part of the sequence (e.g., just the byte “b1”), preventing the model from fully capturing the intended network context. As a result, the mode… view at source ↗

**Figure 2.** A case study on header field length distribution on the ISCX-VPN [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Parameter analysis on field-level masking. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Classification Performance under different sampling datasets. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 16 canonical work pages

[1]

Efficient fine-grained website fingerprinting via encrypted traffic analysis with deep learning,

M. Shen, Z. Gao, L. Zhu, and K. Xu, “Efficient fine-grained website fingerprinting via encrypted traffic analysis with deep learning,” in 2021 IEEE/ACM 29th International Symposium on Quality of Service (IWQOS). IEEE, 2021, pp. 1–10

work page 2021
[2]

Expmd: an explainable framework for traffic identification based on multi-domain features,

L. Yang, Y . Wang, L. Liu, J. Huang, and S. Fu, “Expmd: an explainable framework for traffic identification based on multi-domain features,” in 2024 IEEE/ACM 32nd International Symposium on Quality of Service (IWQoS). IEEE, 2024, pp. 1–10

work page 2024
[3]

Byte-label joint attention learning for packet-grained network traffic classification,

K. Mao, X. Xiao, G. Hu, X. Luo, B. Zhang, and S. Xia, “Byte-label joint attention learning for packet-grained network traffic classification,” in2021 IEEE/ACM 29th International Symposium on Quality of Service (IWQOS). IEEE, 2021, pp. 1–10

work page 2021
[4]

Radd: A real-time and accurate method for ddos detection based on in-network computing,

W. Wang, S. Zhu, Z. Wu, L. Lu, Z. Li, H. Yang, and Y . Zhang, “Radd: A real-time and accurate method for ddos detection based on in-network computing,” inICC 2024-IEEE International Conference on Communications. IEEE, 2024, pp. 3316–3321

work page 2024
[5]

Et-bert: A contextualized datagram representation with pre-training transformers for encrypted traffic classification,

X. Lin, G. Xiong, G. Gou, Z. Li, J. Shi, and J. Yu, “Et-bert: A contextualized datagram representation with pre-training transformers for encrypted traffic classification,” inProceedings of the ACM Web Conference 2022, 2022, pp. 633–642

2022
[6]

A novel self-supervised framework based on masked autoen- coder for traffic classification,

R. Zhao, M. Zhan, X. Deng, F. Li, Y . Wang, Y . Wang, G. Gui, and Z. Xue, “A novel self-supervised framework based on masked autoen- coder for traffic classification,”IEEE/ACM Transactions on Networking, 2024

work page 2024
[7]

Quality of service (qos) in software defined networking (sdn): A survey,

M. Karakus and A. Durresi, “Quality of service (qos) in software defined networking (sdn): A survey,”Journal of Network and Computer Applications, vol. 80, pp. 200–218, 2017

work page 2017
[8]

Accelerating traffic engineer- ing optimization for segment routing: A recommendation perspective,

L. Wang, M. Wang, C. Lin, and Y . Zhang, “Accelerating traffic engineer- ing optimization for segment routing: A recommendation perspective,” Computer Networks, vol. 264, p. 111224, 2025

work page 2025

Show all 25 references

[9]

Flowprint: Semi-supervised mobile-app fingerprinting on encrypted network traf- fic,

T. Van Ede, R. Bortolameotti, A. Continella, J. Ren, D. J. Dubois, M. Lindorfer, D. Choffnes, M. van Steen, and A. Peter, “Flowprint: Semi-supervised mobile-app fingerprinting on encrypted network traf- fic,” inNetwork and distributed system security symposium (NDSS), vol. 27, 2020

2020
[10]

Appscanner: Automatic fingerprinting of smartphone apps from encrypted network traffic,

V . F. Taylor, R. Spolaor, M. Conti, and I. Martinovic, “Appscanner: Automatic fingerprinting of smartphone apps from encrypted network traffic,” in2016 IEEE European Symposium on Security and Privacy (EuroS&P). IEEE, 2016, pp. 439–454

2016
[11]

An analysis of network traffic identification based on decision tree,

J. Dai, Y . Chen, Y . Chen, and A. Meng, “An analysis of network traffic identification based on decision tree,” in2021 International Conference on Artificial Intelligence and Electromechanical Automation (AIEA). IEEE, 2021, pp. 308–311

2021
[12]

On using extreme gradient boosting (xgboost) machine learning algorithm for home network traffic clas- sification,

I. L. Cherif and A. Kortebi, “On using extreme gradient boosting (xgboost) machine learning algorithm for home network traffic clas- sification,” in2019 Wireless Days (WD). IEEE, 2019, pp. 1–6

2019
[13]

Fs-net: A flow sequence network for encrypted traffic classification,

C. Liu, L. He, G. Xiong, Z. Cao, and Z. Li, “Fs-net: A flow sequence network for encrypted traffic classification,” inIEEE INFOCOM 2019- IEEE Conference On Computer Communications. IEEE, 2019, pp. 1171–1179

2019
[14]

Ebsnn: Extended byte segment neural network for network traffic classification,

X. Xiao, W. Xiao, R. Li, X. Luo, H. Zheng, and S. Xia, “Ebsnn: Extended byte segment neural network for network traffic classification,” IEEE Transactions on Dependable and Secure Computing, vol. 19, no. 5, pp. 3521–3538, 2021

2021
[15]

Tfe-gnn: A temporal fusion encoder using graph neural networks for fine-grained encrypted traffic classification,

H. Zhang, L. Yu, X. Xiao, Q. Li, F. Mercaldo, X. Luo, and Q. Liu, “Tfe-gnn: A temporal fusion encoder using graph neural networks for fine-grained encrypted traffic classification,” inProceedings of the ACM Web Conference 2023, 2023, pp. 2066–2075

2023
[16]

Pert: Payload encoding representation from transformer for encrypted traffic classification,

H. Y . He, Z. G. Yang, and X. N. Chen, “Pert: Payload encoding representation from transformer for encrypted traffic classification,” in 2020 ITU Kaleidoscope: Industry-Driven Digital Transformation (ITU K). IEEE, 2020, pp. 1–8

2020
[17]

Netgpt: Generative pretrained transformer for network traffic,

X. Meng, C. Lin, Y . Wang, and Y . Zhang, “Netgpt: Generative pretrained transformer for network traffic,”arXiv preprint arXiv:2304.09513, 2023

2023 arXiv
[18]

A malware- detection method using deep learning to fully extract api sequence features,

S. Zhang, M. Gao, L. Wang, S. Xu, W. Shao, and R. Kuang, “A malware- detection method using deep learning to fully extract api sequence features,”Electronics, vol. 14, no. 1, p. 167, 2025

2025
[19]

A few shots traffic classification with mini-flowpic augmentations,

E. Horowicz, T. Shapira, and Y . Shavitt, “A few shots traffic classification with mini-flowpic augmentations,” inProceedings of the 22nd ACM Internet Measurement Conference, 2022, pp. 647–654

2022
[20]

Realistic website fingerprinting by augmenting network traces,

A. Bahramali, A. Bozorgi, and A. Houmansadr, “Realistic website fingerprinting by augmenting network traces,” inProceedings of the 2023 ACM SIGSAC Conference on Computer and Communications Security, 2023, pp. 1035–1049

2023
[21]

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 Proceedings of the 2nd international conference on information systems security and privacy (ICISSP), 2016, pp. 407–414

2016
[22]

Malware traffic classification using convolutional neural network for representation learning,

W. Wang, M. Zhu, X. Zeng, X. Ye, and Y . Sheng, “Malware traffic classification using convolutional neural network for representation learning,” in2017 International conference on information networking (ICOIN). IEEE, 2017, pp. 712–717

2017
[23]

Towards the development of a realistic multi- dimensional iot profiling dataset,

S. Dadkhah, H. Mahdikhani, P. K. Danso, A. Zohourian, K. A. Truong, and A. A. Ghorbani, “Towards the development of a realistic multi- dimensional iot profiling dataset,” in2022 19th Annual International Conference on Privacy, Security & Trust (PST). IEEE, 2022, pp. 1–11

2022
[24]

Bert: Pre-training of deep bidirectional transformers for language understanding,

J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,”arXiv preprint arXiv:1810.04805, 2018

2018 arXiv
[25]

De- gnn: Dual embedding with graph neural network for fine-grained en- crypted traffic classification,

X. Han, G. Xu, M. Zhang, Z. Yang, Z. Yu, W. Huang, and C. Meng, “De- gnn: Dual embedding with graph neural network for fine-grained en- crypted traffic classification,”Computer Networks, vol. 245, p. 110372, 2024

2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Efficient fine-grained website fingerprinting via encrypted traffic analysis with deep learning,

M. Shen, Z. Gao, L. Zhu, and K. Xu, “Efficient fine-grained website fingerprinting via encrypted traffic analysis with deep learning,” in 2021 IEEE/ACM 29th International Symposium on Quality of Service (IWQOS). IEEE, 2021, pp. 1–10

work page 2021

[2] [2]

Expmd: an explainable framework for traffic identification based on multi-domain features,

L. Yang, Y . Wang, L. Liu, J. Huang, and S. Fu, “Expmd: an explainable framework for traffic identification based on multi-domain features,” in 2024 IEEE/ACM 32nd International Symposium on Quality of Service (IWQoS). IEEE, 2024, pp. 1–10

work page 2024

[3] [3]

Byte-label joint attention learning for packet-grained network traffic classification,

K. Mao, X. Xiao, G. Hu, X. Luo, B. Zhang, and S. Xia, “Byte-label joint attention learning for packet-grained network traffic classification,” in2021 IEEE/ACM 29th International Symposium on Quality of Service (IWQOS). IEEE, 2021, pp. 1–10

work page 2021

[4] [4]

Radd: A real-time and accurate method for ddos detection based on in-network computing,

W. Wang, S. Zhu, Z. Wu, L. Lu, Z. Li, H. Yang, and Y . Zhang, “Radd: A real-time and accurate method for ddos detection based on in-network computing,” inICC 2024-IEEE International Conference on Communications. IEEE, 2024, pp. 3316–3321

work page 2024

[5] [5]

Et-bert: A contextualized datagram representation with pre-training transformers for encrypted traffic classification,

X. Lin, G. Xiong, G. Gou, Z. Li, J. Shi, and J. Yu, “Et-bert: A contextualized datagram representation with pre-training transformers for encrypted traffic classification,” inProceedings of the ACM Web Conference 2022, 2022, pp. 633–642

2022

[6] [6]

A novel self-supervised framework based on masked autoen- coder for traffic classification,

R. Zhao, M. Zhan, X. Deng, F. Li, Y . Wang, Y . Wang, G. Gui, and Z. Xue, “A novel self-supervised framework based on masked autoen- coder for traffic classification,”IEEE/ACM Transactions on Networking, 2024

work page 2024

[7] [7]

Quality of service (qos) in software defined networking (sdn): A survey,

M. Karakus and A. Durresi, “Quality of service (qos) in software defined networking (sdn): A survey,”Journal of Network and Computer Applications, vol. 80, pp. 200–218, 2017

work page 2017

[8] [8]

Accelerating traffic engineer- ing optimization for segment routing: A recommendation perspective,

L. Wang, M. Wang, C. Lin, and Y . Zhang, “Accelerating traffic engineer- ing optimization for segment routing: A recommendation perspective,” Computer Networks, vol. 264, p. 111224, 2025

work page 2025