Pith. sign in

REVIEW 3 major objections 5 minor 3 cited by

A sparse mixture-of-experts transformer pre-trained on unlabeled network flows can match or beat dense pre-trained baselines on traffic security tasks while roughly doubling throughput and halving latency, establishing that model capacity a

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Traffic-MoE is a mixture-of-experts traffic transformer that claims up to 12.38% better detection Macro-F1 than dense pre-trained baselines while roughly doubling throughput and cutting latency by roughly half.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection Traffic-MoE is a credible, well-evaluated first application of sparse MoE to traffic transformers, but the headline efficiency numbers rest on an underspecified dense comparison and no variance reporting; worth refereeing, not worth citing yet. the 3 major comments →

arxiv 2601.00357 v2 pith:UFTE4ER2 submitted 2026-01-01 cs.CR

Traffic-MoE: A Sparse Foundation Model for Network Traffic Security Analysis

Jiajun Zhou , Changhui Sun , Wentao Fu , Meng Shen , Shanqing Yu , Qi Xuan This is my paper
classification cs.CR
keywords mixture-of-expertsnetwork traffic classificationencrypted trafficpre-trained foundation modelfew-shot learningTor/VPN detectionintrusion detectiontraffic tokenization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Traffic-MoE is a pre-trained transformer for network security analysis that replaces dense feed-forward layers with a sparse mixture-of-experts (MoE) layer, routing each token to only two of many specialized experts. The paper claims this sparsity decouples model capacity from compute: the model matches or beats the strongest dense pre-trained baselines on six traffic classification tasks (Tor service classification, VPN service classification, IoT and IoMT attack detection) while roughly doubling throughput, halving latency, and cutting peak GPU memory by a third. It further claims the sparse experts act as a regularizer, preserving detection quality under few-shot supervision and resisting traffic-shaping and distribution shifts that degrade dense baselines. The central mechanism is Traffic2Token, which serializes packet flows into byte-bigram tokens with metadata markers, on which the model learns next-token prediction over ~2M unlabeled flows before task-specific fine-tuning.

Core claim

Traffic-MoE establishes that a sparsely activated transformer can be both a larger-capacity and a more deployable network-traffic classifier than a dense one. After pre-training by next-token prediction on a corpus of over two million unlabeled flows spanning malware, intrusion, IoT, and VPN/Tor traffic, the model is fine-tuned on four downstream datasets via temporal slicing, mean pooling, and layer-wise learning-rate decay. In the paper's evaluation it reaches 0.89 Macro-F1 on Tor traffic, 0.8332 on mixed VPN traffic, 0.8839 on IoMT attacks, and 0.7824 on IoT attacks, numbers above the compared baselines; at batch size 64 it reports 186.25 samples/sec throughput versus 131.56 for its dense

What carries the argument

The central object is the top-2 sparse mixture-of-experts layer: a bank of N SwiGLU-based experts plus one always-on shared expert, with a gating network that selects the two most relevant experts per token. Because only k experts plus the shared expert are activated, the model's parameter count grows with N while its inference cost stays tied to k, so capacity and latency are decoupled. A load-balancing auxiliary loss keeps the router from collapsing onto a few experts, and the shared expert absorbs protocol-universal patterns so that specialized experts can specialize in distinct traffic behaviors. The supporting tokenizer, Traffic2Token, maps each packet's metadata bytes and a sampled pay

Load-bearing premise

The load-bearing premise is that the dense comparison model really is the same size as Traffic-MoE, and that the pre-training corpus does not silently overlap with the labels or distributions used in the downstream tasks.

What would settle it

Recompute the headline comparison with a dense network of exactly the same layer count, hidden width, and total parameters as Traffic-MoE, and check whether the throughput, latency, and memory advantages shrink; then remove every pre-training flow that also appears in the fine-tuning test sets and see whether the few-shot and distribution-shift margins survive.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the efficiency figures hold at deployment scale, a foundation-model-level traffic classifier can run on the same hardware that currently runs smaller dense models, removing a stated obstacle to encrypted-traffic detection at network gateways.
  • The reported few-shot results imply that a sparse router pre-trained on unlabeled flows can match the full-supervision accuracy of dense baselines with 10-20% of the labels, which matters for zero-day and long-tail attack classes.
  • The robustness results under time, proportion, and composition shifts indicate the model relies on protocol grammar rather than dataset-specific statistics, so it may degrade more gracefully when traffic patterns drift.
  • The Tor-specific gains suggest that routing, not just additional parameters, is what recovers signal from traffic shaping: dense baselines trained identically on the same corpus do not show the same resilience.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If expert specialization is the true cause of the Tor/VPN gains, then routing visualizations such as Figure 8 could be turned into a diagnostic: clusters of flows that consistently activate the same experts should correspond to semantically related applications, and adversarial flows should be detectable by anomalous routing patterns.
  • The paper leaves open whether the shared expert alone, rather than the mix of shared plus specialized experts, could carry the whole accuracy gain; an ablation that replaces specialized experts with a single large shared expert would isolate the contribution of routing.
  • A natural stress test is to pre-train on flows deliberately disjoint from the downstream evaluation sets; if the few-shot and OOD gains shrink, the benefits are partly dataset overlap rather than general traffic understanding.
  • The efficiency comparison would generalize more convincingly if repeated on CPU and edge hardware, since real gateways often lack A100-class GPUs; the reported GPU numbers may not transfer directly.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes Traffic-MoE, a sparse Mixture-of-Experts transformer for network traffic security analysis. Input flows are serialized into byte-level token sequences by a new module called Traffic2Token, then processed by a causal transformer in which dense feed-forward layers are replaced by a shared expert plus top-2 routed specialized experts. The model is pre-trained with a next-token prediction objective plus a load-balancing auxiliary loss on a corpus of roughly two million flows drawn from five public datasets, and fine-tuned on six downstream tasks spanning Tor/NonTor service classification, VPN/NonVPN service classification, and IoT/IoMT attack detection. The paper reports state-of-the-art or competitive Macro-F1 on those tasks, substantial inference-efficiency gains over dense baselines and prior pre-trained models, superior few-shot behavior, and robustness to three types of distribution shift.

Significance. If the reported results are correct, the paper makes a useful contribution: it demonstrates that sparse expert activation can be carried over to network traffic analysis and that a large-capacity model can remain efficient enough for throughput-sensitive deployment. The evaluation is broad — six tasks, four public datasets, six baselines, ablations for MoE-vs-dense, pre-training, load-balancing loss, input representations, top-k, payload length, packet count, few-shot curves, and three OOD shift types — and it compares against independently published methods on public benchmarks. The central claim, however, rests on two controls that are not documented: the claimed parameter-matched dense twin, and the avoidance of train/test leakage when pre-training and fine-tuning draw from the same datasets. Both are load-bearing for the headline efficiency and few-shot/OOD conclusions.

major comments (3)
  1. [§5.2.2, Table 4, §5.7, Appendix E.3.1] The 'MoE→Dense' ablation is called parameter-matched, but no model-size information is reported anywhere: the number of layers L, hidden dimension d, expert count N, expert intermediate dimension d'/k, total parameter count, active parameter count, or FLOPs are absent. Without these numbers the efficiency claims in §5.7 — 41.57% throughput gain, 29.36% latency reduction, and roughly 40% computation bypass versus the dense variant — are not interpretable. The accuracy comparison in Table 4 is also fragile: the dense variant wins on CICIoMT2024 (0.8946 vs 0.8839) and CICIoT2023 (0.7894 vs 0.7824), ties on NonTor (0.8904 vs 0.8900), and the remaining gains are 0.8–3.0 M-F1 points with no error bars or seed variance. Please report the full architecture table and repeat the dense/MoE comparison over at least three seeds.
  2. [§5.1.1, Table 1, §5.1.2] The pre-training corpus includes CICIoT2023, CICIoMT2024, and ISCXVPN2016(NonVPN), and the same three datasets are later used for fine-tuning. The paper says the pre-training flows are unlabeled, but unlabeled flows from the same source still carry distributional information about the downstream test sets, especially after the 8:1:1 split is applied within each dataset. This threatens the few-shot and OOD robustness conclusions: the model may be evaluated on flows whose source distribution was already seen during pre-training. Please either (a) exclude the downstream datasets from pre-training and re-run the key experiments, or (b) demonstrate that the pre-training and fine-tuning session splits are disjoint and quantify the performance change when the overlap is removed.
  3. [Abstract vs. §5.7] The efficiency numbers are inconsistent across versions of the paper. The supplied abstract reports a 70.42% throughput increase and 41.39% latency reduction; the full-text abstract and the contributions list report 91.62% throughput and 47.81% latency; §5.7 reports 41.57% throughput and 29.36% latency against the dense variant and 91.62% throughput with a 38.72% memory saving against NetGPT. These are not the same comparisons, and the abstract's numbers are not derivable from the body. The detection improvement claim of 'up to 12.38%' is also not tied to a specific row of Table 2. Please harmonize the abstract, contributions, and Section 5.7, and state precisely which baseline each efficiency percentage refers to.
minor comments (5)
  1. [Table 2] The ET-BERT row for CICIoMT2024 appears to contain an anomalous 0.9769 in the accuracy column before recycling into the M-PR/M-RC/M-F1 columns. Please verify the table formatting.
  2. [§5.9 / Table 6] The definitions of Time-shift, Proportion-shift, and Compose-shift are deferred to Appendix E.6. Since robustness under distribution shift is a central claim, at least a one-sentence definition of each shift should appear in the main text, and the appendix must be included in the submission.
  3. [§5.6.1] The phrase 'parameter-matched dense baseline' is used twice without any supporting parameter count. If the authors intend a fixed total-parameter match, this should be stated explicitly in the text and not only in the appendix.
  4. [§5.4.2 / Appendix E.2] The explanation of Audio/Email degradation under Tor is reasonable, but it is post-hoc and not tested. A short experiment restricting the input window or ablating padding could make the claim more than an interpretation.
  5. [Conclusion] The conclusion repeats the 91.62% / 47.81% / 38.72% numbers without specifying the comparison baseline; it should match the clarified Section 5.7 figures.

Circularity Check

0 steps flagged

No significant circularity: Traffic-MoE's core claims rest on external benchmarks, measured efficiency numbers, and author-trained ablation controls; the flagged issues are reporting/leakage concerns, not equivalence-by-construction.

full rationale

The paper's derivation chain is self-contained in the sense required by the circularity test. Traffic2Token produces token sequences from packet metadata/payload bytes; pre-training is next-token prediction (Eqs. 10-12) plus a load-balancing auxiliary loss, neither of which consumes downstream labels; fine-tuning is standard cross-entropy over public benchmark labels (Eqs. 13-15). Downstream performance is compared against independently published baselines (ET-BERT, NetGPT, TrafficFormer, etc.) on public datasets, and the MoE-vs-dense and pre-training ablations are author-trained controls rather than fitted values renamed as predictions. The efficiency claims are empirical throughput/latency/memory measurements (Sec. 5.7, Fig. 7), not theorem-level consequences of the architecture alone; the sparse-activation speedup follows from top-2 routing by design, but the reported percentages are measured. The main concerns raised in the paper — the under-specified 'parameter-matched dense baseline' (Table 4, Sec. 5.6.1) with no layer count, expert count, hidden dim, or parameter totals, and the overlap between pre-training corpora and fine-tuning datasets (Table 1) — are correctness, reproducibility, and potential leakage issues, not circular reasoning: no equation in the paper reduces a predicted quantity to its own input, and no load-bearing argument is supported by a self-citation chain. The self-citations (e.g., refs. [42], [58]) appear only in related-work/motivation and are not used to justify the central result. Accordingly, the circularity score is 0.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

The paper deliberately uses very few hand-tuned knobs beyond standard MoE machinery, but the central free parameter is the k=2 expert count, which is tuned on the eval tasks. The pre-train/fine-tune dataset overlap is the main unstated assumption; the 'expert specialisation' story is the main invented-entity risk. No truly new physical or mathematical entities are introduced.

free parameters (5)
  • k (number of active experts) = 2
    Chosen by scanning k in {1,2,3,4,6} over six tasks (Figure 11) without reporting variance; k=2 is the tuned peak, and many claims inherit the benefits of this tuned choice.
  • Auxiliary loss weight lambda_aux = 0.02
    Set by hand, only one value reported; the w/o-Laux ablation shows the model is sensitive to this term.
  • Number of experts N and layer count L = N/A
    Never reported in the provided text, so the efficiency and capacity claims cannot be checked against actual parameter counts.
  • J (payload length) and K (packet count) = J=40, K=10
    Chosen to 'balance computational efficiency and early-stage detection' (Sec 5.2.2) and then swept in ablations; the specific values are tunings, not derived.
  • LLRD decay factor xi and base LR eta0 = xi=0.9, eta0=5e-5
    Standard fine-tuning hyperparameters, set by hand.
axioms (4)
  • domain assumption MoE top-2 routing plus load-balancing loss confers domain specialization that improves generalization on obfuscated traffic
    The whole paper's explanatory story (Sec 5.3, 5.8) assumes expert specialization is the mechanism behind Tor/VPN gains; the only evidence is the MoE-vs-dense ablation and activation heatmaps, both of which also admit a simpler explanation: extra parameters or regularizing sparsity.
  • domain assumption Pre-training on the five listed public datasets plus fine-tuning on overlapping dataset families does not leak label-distribution information into downstream evaluation
    Sec 5.1 pre-trains on USTC-TFC2016, UNSW-NB15, CICIoT2023, CICIoMT2024, ISCXVPN2016-NonVPN and fine-tunes on CICIoT2023, CICIoMT2024, ISCXVPN2016, ISCXTor2016; three of the four fine-tuning benchmarks have the same origin as pre-training data. Scrutability of the few-shot/OOD gains depends on this assumption.
  • domain assumption The first K=10 packets of a flow carry enough behavioral signal to classify the whole flow
    Used throughout the tokenizer (Sec 4.1.1) and tested only implicitly in the packet-count ablation; for long-lived flows (torrent, video streams) this truncation discards most of the session.
  • domain assumption Byte bigram + packet metadata tokenization is a sufficient representation for security-relevant semantics
    The Traffic2Token design (Sec 4.1) is asserted, not derived; the ablations 'Header Only' and 'Payload Only' show both modalities matter but do not justify the specific concatenation/sampling.
invented entities (2)
  • Traffic2Token byte-level token vocabulary with [PD]/[PY]/[PAD]/[END]/[UNK] markers no independent evidence
    purpose: Convert raw flows into token sequences for pre-training; markers preserve packet boundaries
    A design artefact, not an empirical discovery; its value is demonstrated only through the model's downstream performance, which is exactly the thing it is used to explain.
  • Specialized experts as interpretable 'domain experts' that decouple obfuscation artifacts no independent evidence
    purpose: Supports the claim that sparse routing is the causal mechanism for accuracy-efficiency gains
    Fig 8 shows activation heatmaps, but no test shows that routing to a particular expert is either necessary or sufficient for the Tor/VPN gains; this is a post-hoc interpretation.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Traffic-MoE: A Sparse Foundation Model for Network Traffic Security Analysis." pith.science (2026). https://pith.science/paper/UFTE4ER2

@misc{pith2026260100357,
  author       = {Pith},
  title        = {Pith review of: Traffic-MoE: A Sparse Foundation Model for Network Traffic Security Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UFTE4ER2}},
  note         = {Machine review of arXiv:2601.00357}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

As adversaries increasingly weaponize encryption and protocol obfuscation to evade traffic detection, traditional methods are rendered obsolete, necessitating deep learning to unmask sophisticated threats. However, the prohibitive computational costs of existing large models create a critical defense gap, hindering their deployment in real-time and throughput-sensitive environments. To close this vulnerability, we introduce Traffic-MoE, a sparse foundation model tailored for traffic security analysis. By dynamically routing traffic tokens to a small subset of specialized experts, Traffic-MoE effectively decouples model capacity from computational overhead. Extensive evaluations across four security-oriented tasks demonstrate that Traffic-MoE achieves state-of-the-art or highly competitive performance compared to leading competitors. Crucially, it delivers a 70.42% increase in throughput, reduces inference latency by 41.39% while significantly optimizing GPU memory consumption. Beyond efficiency, Traffic-MoE exhibits superior robustness against adversarial traffic shaping and maintains strong detection capabilities in few-shot scenarios, establishing a scalable and resilient paradigm for modern network traffic security analysis.

Figures

Figures reproduced from arXiv: 2601.00357 by Changhui Sun, Jiajun Zhou, Meng Shen, Qi Xuan, Shanqing Yu, Wentao Fu.

Figure 1
Figure 1. Figure 1: Threat model and operational scenario. specialized deep inspection is reserved for ambiguous, high-risk, or encrypted flows mirrored to the analysis engine via out-of-path inspection. This design ensures zero impact on network latency while allowing detection results to trigger dynamic policy updates (e.g., blocking or QoS adjustment) in real-time [7, 19]. Crucially, given the immense volume of backbone tr… view at source ↗
Figure 2
Figure 2. Figure 2: The overall framework of Traffic-MoE. It consists of the Traffic2Token serialization module, the sparse Mixture-of￾Experts (MoE) backbone, and the pre-training/fine-tuning pipeline for downstream security tasks. Packet Tokens [PD] Packet Features Tokens [PY] Payload Tokens Flow Token ID Sequence Packet Tokens … [END] Token Packet Length Transmission Direction TCP Flag Bits Inter-arrival Time IP Protocol Pa… view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of flow token sequence representation. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of Traffic-MoE backbone architecture. 4.2.1 Causal Masked Self-Attention. Network flows evolve through ordered protocol phases such as handshake exchanges, tunneling encapsulation, retransmissions, and periodic keep-alives, which manifest as temporal structures in their byte-level token sequences. To model these sequential dependencies while respecting real-world monitoring constraints where f… view at source ↗
Figure 5
Figure 5. Figure 5: Performance confusion matrices of Traffic-MoE across different service classification tasks [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Performance evaluation in few-shot supervision. [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Inference efficiency comparison among different ThroughputAverage LatencyPeak Memory Consumption [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Expert Activation Across Model Layers. Method 5% 10% 20% 40% 100% AppScann 0.9117 0.9263 0.9337 0.9394 0.9468 Flowprint 0.752 0.457 0.9031 0.8757 0.9213 FS-Net 0.8583 0.8457 0.891 0.9135 0.9406 ET-BERT 0.9345 0.9149 0.9465 0.9517 0.9653 NetGPT 0.9173 0.9219 0.9413 0.9477 0.9655 Trafficfor 0.9213 0.9336 0.9195 0.9325 0.9554 TrafficMo 0.9724 0.9688 0.976 0.9793 0.9827 Method 5% 10% 20% 40% 100% AppScann 0.67… view at source ↗
Figure 9
Figure 9. Figure 9: Performance under different distribution shifts. [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Performance confusion matrices of Traffic-MoE across different traffic downstream tasks. patterns (which persist in P2P/VoIP), proving robustness against traffic shaping where semantic structure remains intact. E.3 Detailed Ablation Analysis To quantify the contribution of key model components and verify the rationality of the architectural design, we conduct a multi￾dimensional ablation analysis based on… view at source ↗
Figure 11
Figure 11. Figure 11: Impact of Top-𝑘 parameters on model perfor￾mance. Method 5% 10% 20% 40% 100% AppScann 0.9117 0.9263 0.9337 0.9394 0.9468 Flowprint 0.752 0.457 0.9031 0.8757 0.9213 FS-Net 0.8583 0.8457 0.891 0.9135 0.9406 ET-BERT 0.9345 0.9149 0.9465 0.9517 0.9653 NetGPT 0.9173 0.9219 0.9413 0.9477 0.9655 Trafficfor 0.9213 0.9336 0.9195 0.9325 0.9554 TrafficMo 0.9724 0.9688 0.976 0.9793 0.9827 Method 5% 10% 20% 40% 100% A… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. NetVAD: Foundation-Model Representation Learning for Identifier-Free Unsupervised Intrusion Detection

    cs.CR 2026-05 unverdicted novelty 5.0

    NetVAD is a strictly identifier-free VAE that projects frozen foundation-model network representations into a latent space trained solely on benign traffic, reporting 98% micro F1 and 96% macro F1 on ToN-IoT.

  2. NetVAD: Foundation-Model Representation Learning for Identifier-Free Unsupervised Intrusion Detection

    cs.CR 2026-05 unverdicted novelty 5.0

    NetVAD uses a strictly identifier-free VAE on frozen foundation model embeddings, trained solely on benign traffic, to achieve 98% micro F1 and 96% macro F1 on ToN-IoT for unsupervised intrusion detection.

  3. Towards Intrusion Detection Systems for RPL-based IoT Networks using Foundation Models

    cs.CR 2026-06 unverdicted novelty 4.0

    Fine-tuning the MOMENT foundation model on simulated RPL statistics enables multi-class attack identification with performance comparable to state-of-the-art methods.

Reference graph

Works this paper leans on

71 extracted references · 4 linked inside Pith · cited by 2 Pith papers

  1. [1]

    Mohammed Ali Al-Garadi, Amr Mohamed, Abdulla Khalid Al-Ali, Xiaojiang Du, Ihsan Ali, and Mohsen Guizani. 2020. A survey of machine and deep learning methods for internet of things (IoT) security.IEEE communications surveys & tutorials22, 3 (2020), 1646–1685

    2020

  2. [2]

    Blake Anderson and David McGrew. 2016. Identifying encrypted malware traffic with contextual flow data. InProceedings of the 2016 ACM workshop on artificial intelligence and security. 35–46

    2016

  3. [3]

    Blake Anderson, Subharthi Paul, and David McGrew. 2018. Deciphering mal- ware’s use of TLS (without decryption).Journal of Computer Virology and Hacking Techniques14, 3 (2018), 195–211

    2018

  4. [4]

    Manos Antonakakis, Tim April, Michael Bailey, Matt Bernhard, Elie Bursztein, Jaime Cochran, Zakir Durumeric, J Alex Halderman, Luca Invernizzi, Michalis Kallitsis, et al. 2017. Understanding the mirai botnet. In26th USENIX security symposium (USENIX Security 17). 1093–1110

    2017

  5. [5]

    Ahmad Azab, Mahmoud Khasawneh, Saed Alrabaee, Kim-Kwang Raymond Choo, and Maysa Sarsour. 2024. Network traffic classification: Techniques, datasets, and challenges.Digital Communications and Networks10, 3 (2024), 676–692

    2024

  6. [6]

    Alireza Bahramali, Ardavan Bozorgi, and Amir Houmansadr. 2023. Realistic website fingerprinting by augmenting network traces. InProceedings of the 2023 ACM SIGSAC Conference on Computer and Communications Security. 1035–1049

    2023

  7. [7]

    Juan Camilo Correa Chica, Jenny Cuatindioy Imbachi, and Juan Felipe Botero Vega. 2020. Security in SDN: A comprehensive survey.Journal of Network and Computer Applications159 (2020), 102595

    2020

  8. [8]

    Sajjad Dadkhah, Euclides Carlos Pinto Neto, Raphael Ferreira, Reginald Chuk- wuka Molokwu, Somayeh Sadeghi, and Ali A Ghorbani. 2024. CICIoMT2024: A benchmark dataset for multi-protocol security assessment in IoMT.Internet of Things28 (2024), 101351

    2024

  9. [9]

    Roger Dingledine, Nick Mathewson, and Paul Syverson. 2004. Tor: The second- generation onion router. (2004)

    2004

  10. [10]

    William Fedus, Barret Zoph, and Noam Shazeer. 2022. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity.Journal of Machine Learning Research23, 120 (2022), 1–39

    2022

  11. [11]

    Igor Fosić, Drago Žagar, Krešimir Grgić, and Višnja Križanović. 2023. Anom- aly detection in NetFlow network traffic using supervised machine learning algorithms.Journal of industrial information integration33 (2023), 100466

    2023

  12. [12]

    Chuanpu Fu, Qi Li, Meng Shen, and Ke Xu. 2024. Detecting tunneled flooding traffic via deep semantic analysis of packet length patterns. InProceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security. 3659–3673

    2024

  13. [13]

    Gerard Drapper Gil, Arash Habibi Lashkari, Mohammad Mamun, and Ali A Ghorbani. 2016. Characterization of encrypted and VPN traffic using time-related features. InProceedings of the 2nd international conference on information systems security and privacy (ICISSP 2016). SciTePress Setúbal, Portugal, 407–414

    2016

  14. [14]

    Jie Gui, Tuo Chen, Jing Zhang, Qiong Cao, Zhenan Sun, Hao Luo, and Dacheng Tao. 2024. A survey on self-supervised learning: Algorithms, applications, and future trends.IEEE Transactions on Pattern Analysis and Machine Intelligence46, 12 (2024), 9052–9071

    2024

  15. [15]

    Chaoqun Guo, Nan Wang, Yuanlin Sun, and Dalin Zhang. 2023. DTC: Addressing the long-tailed problem in intrusion detection through the divide-then-conquer paradigm. In2023 IEEE 29th International Conference on Parallel and Distributed Systems (ICPADS). IEEE, 1319–1326

    2023

  16. [16]

    Hong Ye He, Zhi Guo Yang, and Xiang Ning Chen. 2020. PERT: Payload encoding representation from transformer for encrypted traffic classification. In2020 ITU Kaleidoscope: Industry-Driven Digital Transformation (ITU K). IEEE, 1–8

    2020

  17. [17]

    Jordan Holland, Paul Schmitt, Nick Feamster, and Prateek Mittal. 2021. New directions in automated traffic analysis. InProceedings of the 2021 ACM SIGSAC conference on computer and communications security. 3366–3383

    2021

  18. [18]

    Amir Houmansadr, Chad Brubaker, and Vitaly Shmatikov. 2013. The parrot is dead: Observing unobservable network communications. In2013 IEEE Symposium on Security and Privacy. IEEE, 65–79

    2013

  19. [19]

    Hongxin Hu, Wonkyu Han, Gail-Joon Ahn, and Ziming Zhao. 2014. FLOW- GUARD: Building robust firewalls for software-defined networks. InProceedings of the third workshop on Hot topics in software defined networking. 97–102

    2014

  20. [20]

    Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al . 2024. Mixtral of experts.arXiv preprint arXiv:2401.04088(2024)

    Pith/arXiv arXiv 2024

  21. [21]

    Adam Langley, Alistair Riddoch, Alyssa Wilk, Antonio Vicente, Charles Krasic, Dan Zhang, Fan Yang, Fedor Kouranov, Ian Swett, Janardhan Iyengar, et al. 2017. The quic transport protocol: Design and internet-scale deployment. InProceedings of the conference of the ACM special interest group on data communication. 183– 196

    2017

  22. [22]

    Arash Habibi Lashkari, Gerard Draper Gil, Mohammad Saiful Islam Mamun, and Ali A Ghorbani. 2017. Characterization of tor traffic using time based features. InInternational conference on information systems security and privacy, Vol. 2. SciTePress, 253–262

    2017

  23. [23]

    Hyunwoo Lee, Doowon Kim, and Yonghwi Kwon. 2021. TLS 1.3 in practice: How TLS 1.3 contributes to the internet. InProceedings of the Web Conference 2021. 70–79

    2021

  24. [24]

    Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. 2020. Gshard: Scaling giant models with conditional computation and automatic sharding.arXiv preprint arXiv:2006.16668(2020)

    Pith/arXiv arXiv 2020

  25. [25]

    Peiyang Li, Ye Wang, Qi Li, Zhuotao Liu, Ke Xu, Ju Ren, Zhiying Liu, and Ruilin Lin. 2023. Learning from limited heterogeneous training data: Meta-learning for unsupervised zero-day web attack detection across web domains. InProceedings of the 2023 ACM SIGSAC Conference on Computer and Communications Security. 1020–1034

    2023

  26. [26]

    Xinjie Lin, Gang Xiong, Gaopeng Gou, Zhen Li, Junzheng Shi, and Jing Yu

  27. [27]

    Chang Liu, Longtao He, Gang Xiong, Zigang Cao, and Zhen Li. 2019. FS-Net: A flow sequence network for encrypted traffic classification. InIEEE INFOCOM 2019-IEEE Conference On Computer Communications. IEEE, 1171–1179

    2019

  28. [28]

    Ya Liu, Xiao Wang, Bo Qu, and Fengyu Zhao. 2024. ATVITSC: A Novel Encrypted Traffic Classification Method Based on Deep Learning.IEEE Transactions on Information Forensics and Security19 (2024), 9374–9389

    2024

  29. [29]

    Mohammad Lotfollahi, Mahdi Jafari Siavoshani, Ramin Shirali Hossein Zade, and Mohammdsadegh Saberian. 2020. Deep packet: A novel approach for encrypted traffic classification using deep learning.Soft Computing24, 3 (2020), 1999–2012

    2020

  30. [30]

    Xuying Meng, Chungang Lin, Yequan Wang, and Yujun Zhang. 2023. Net- GPT: Generative pretrained transformer for network traffic.arXiv preprint arXiv:2304.09513(2023)

    Pith/arXiv arXiv 2023

  31. [31]

    Andrew W Moore and Konstantina Papagiannaki. 2005. Toward the accurate identification of network applications. InInternational workshop on passive and active network measurement. Springer, 41–54

    2005

  32. [32]

    Nour Moustafa and Jill Slay. 2015. UNSW-NB15: a comprehensive data set for network intrusion detection systems (UNSW-NB15 network data set). In2015 military communications and information systems conference (MilCIS). IEEE, 1–6

    2015

  33. [33]

    Euclides Carlos Pinto Neto, Sajjad Dadkhah, Raphael Ferreira, Alireza Zohourian, Rongxing Lu, and Ali A Ghorbani. 2023. CICIoT2023: A real-time dataset and benchmark for large-scale attacks in IoT environment.Sensors23, 13 (2023), 5941

    2023

  34. [34]

    Eva Papadogiannaki and Sotiris Ioannidis. 2021. A survey on encrypted network traffic analysis applications, techniques, and countermeasures.ACM Computing Surveys (CSUR)54, 6 (2021), 1–35

    2021

  35. [35]

    Nicolas Papernot, Patrick McDaniel, Arunesh Sinha, and Michael P Wellman

  36. [36]

    2018.The transport layer security (TLS) protocol version 1.3

    Eric Rescorla. 2018.The transport layer security (TLS) protocol version 1.3. Techni- cal Report

    2018

  37. [37]

    Martin Roesch et al. 1999. Snort: Lightweight intrusion detection for networks.. InLisa, Vol. 99. 229–238. CCS’25, October 13-17, 2025, Taipei, ON, Taiwan Chen et al

    1999

  38. [38]

    Matthew Roughan, Subhabrata Sen, Oliver Spatscheck, and Nick Duffield. 2004. Class-of-service mapping for QoS: a statistical signature-based approach to IP traffic classification. InProceedings of the 4th ACM SIGCOMM conference on Internet measurement. 135–148

    2004

  39. [39]

    Abid Saber, Belkacem Fergani, and Moncef Abbas. 2018. Encrypted traffic classifi- cation: Combining over-and under-sampling through a pca-svm. InInternational conference on pattern analysis and intelligent systems (PAIS). IEEE, 1–5

    2018

  40. [40]

    Yam Sharon, David Berend, Yang Liu, Asaf Shabtai, and Yuval Elovici. 2022. Tantra: Timing-based adversarial network traffic reshaping attack.IEEE Transactions on Information Forensics and Security17 (2022), 3225–3237

    2022

  41. [41]

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. 2017. Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer. InInternational Conference on Learning Representations. 1–19

    2017

  42. [42]

    Meng Shen, Jinhe Wu, Ke Ye, Ke Xu, Gang Xiong, and Liehuang Zhu. 2025. Robust Detection of Malicious Encrypted Traffic via Contrastive Learning.IEEE Transactions on Information Forensics and Security(2025)

    2025

  43. [43]

    Meng Shen, Ke Ye, Xingtong Liu, Liehuang Zhu, Jiawen Kang, Shui Yu, Qi Li, and Ke Xu. 2022. Machine learning-powered encrypted network traffic analysis: A comprehensive survey.IEEE Communications Surveys & Tutorials25, 1 (2022), 791–824

    2022

  44. [44]

    Robin Sommer and Vern Paxson. 2010. Outside the closed world: On using machine learning for network intrusion detection. In2010 IEEE symposium on security and privacy. IEEE, 305–316

    2010

  45. [45]

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. 2024. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing568 (2024), 127063

    2024

  46. [46]

    Vincent F Taylor, Riccardo Spolaor, Mauro Conti, and Ivan Martinovic. 2017. Robust smartphone app identification via encrypted network traffic analysis. IEEE Transactions on Information Forensics and Security13, 1 (2017), 63–78

    2017

  47. [47]

    Thijs Van Ede, Riccardo Bortolameotti, Andrea Continella, Jingjing Ren, Daniel J Dubois, Martina Lindorfer, David Choffnes, Maarten Van Steen, and Andreas Peter. 2020. Flowprint: Semi-supervised mobile-app fingerprinting on encrypted network traffic. InNetwork and distributed system security symposium (NDSS), Vol. 27. 1–18

    2020

  48. [48]

    Wei Wang, Ming Zhu, Xuewen Zeng, Xiaozhou Ye, and Yiqiang Sheng. 2017. Malware traffic classification using convolutional neural network for representa- tion learning. In2017 International conference on information networking (ICOIN). IEEE, 712–717

    2017

  49. [49]

    Guanglu Wei and Zhonghua Wang. 2021. Adoption and realization of deep learning in network traffic anomaly detection device design.Soft Computing-A Fusion of Foundations, Methodologies & Applications25, 2 (2021), 1147–1158

    2021

  50. [50]

    Yonghui Wu. 2016. Google’s neural machine translation system: Bridging the gap between human and machine translation.arXiv preprint arXiv:1609.08144 (2016)

    Pith/arXiv arXiv 2016

  51. [51]

    Diwen Xue, Reethika Ramesh, Arham Jain, Michaelis Kallitsis, J Alex Halder- man, Jedidiah R Crandall, and Roya Ensafi. 2025. OpenVPN is open to VPN fingerprinting.Commun. ACM68, 1 (2025), 79–87

    2025

  52. [52]

    2019.𝐷𝑒𝑒𝑝−𝐹𝑢𝑙𝑙−𝑅𝑎𝑛𝑔𝑒 : a deep learning based network encrypted traffic classification and intrusion detection framework.IEEE Access7 (2019), 45182–45190

    Yi Zeng, Huaxi Gu, Wenting Wei, and Yantao Guo. 2019.𝐷𝑒𝑒𝑝−𝐹𝑢𝑙𝑙−𝑅𝑎𝑛𝑔𝑒 : a deep learning based network encrypted traffic classification and intrusion detection framework.IEEE Access7 (2019), 45182–45190

    2019

  53. [53]

    Biao Zhang and Rico Sennrich. 2019. Root mean square layer normalization. Advances in neural information processing systems32 (2019)

    2019

  54. [54]

    Haozhen Zhang, Le Yu, Xi Xiao, Qing Li, Francesco Mercaldo, Xiapu Luo, and Qixu Liu. 2023. TFE-GNN: A temporal fusion encoder using graph neural networks for fine-grained encrypted traffic classification. InProceedings of the ACM web conference. 2066–2075

    2023

  55. [55]

    Menghao Zhang, Guanyu Li, Shicheng Wang, Chang Liu, Ang Chen, Hongxin Hu, Guofei Gu, Qianqian Li, Mingwei Xu, and Jianping Wu. 2020. Poseidon: Mitigating volumetric ddos attacks with programmable switches. Inthe 27th Network and Distributed System Security Symposium (NDSS 2020)

    2020

  56. [56]

    Ruijie Zhao, Mingwei Zhan, Xianwen Deng, Yanhao Wang, Yijun Wang, Guan Gui, and Zhi Xue. 2023. Yet another traffic classifier: A masked autoencoder based traffic transformer with multi-level flow representation. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 37. 5420–5427

    2023

  57. [57]

    Guangmeng Zhou, Xiongwen Guo, Zhuotao Liu, Tong Li, Qi Li, and Ke Xu

  58. [58]

    Jiajun Zhou, Wentao Fu, Hao Song, Shanqing Yu, and Qi Xuan. 2025. Multiview Correlation-Aware Network Traffic Detection on Flow Hypergraph.IEEE Internet of Things Journal12, 24 (2025), 55922–55935

    2025

  59. [59]

    Shitong Zhu, Shasha Li, Zhongjie Wang, Xun Chen, Zhiyun Qian, Srikanth V Krishnamurthy, Kevin S Chan, and Ananthram Swami. 2020. You do (not) be- long here: detecting DPI evasion attacks with context learning. InProceedings of the 16th International Conference on emerging Networking EXperiments and Technologies. 183–197. Traffic-MoE: A Sparse Foundation M...

    2020

  60. [63]

    The task covers 33 distinct attack behaviors categorized into 7 major classes, including DDoS, Brute Force, Web attacks, and Mirai botnets

    IoT Attack Detection.Based on theCICIoT2023[ 33] dataset, this scenario simulates a complex network environment consisting of 105 real IoT devices. The task covers 33 distinct attack behaviors categorized into 7 major classes, including DDoS, Brute Force, Web attacks, and Mirai botnets. This task evaluates the model’s capability for fine-grained classific...

  61. [64]

    IoMT Attack Detection.Based on theCICIoMT2024[ 8] dataset, the data originates from a testbed containing 40 real/sim- ulated medical devices, covering 18 targeted attacks across vari- ous protocols such as Wi-Fi and MQTT. Given the high sensitiv- ity of healthcare environments, this task focuses on examining the model’s detection performance in specialize...

  62. [65]

    NonVPNService Classification

    VPN/NonVPN Service Classification.Based on theIS- CXVPN2016[ 13] dataset. This task is highly challenging as VPN technologies introduce multi-layer encryption and tunnel encapsu- lation, significantly altering traffic statistical characteristics. We con- struct two sub-tasks: “NonVPNService Classification” and “Mixed Traffic (VPNandNonVPN) Service Classif...

  63. [66]

    Tor Traffic

    Tor/NonTor Anonymous Service Classification.Based on theISCXTor2016[ 22] dataset. The Onion Router (Tor) achieves high anonymity through multi-layer encryption and node hopping. We divide the dataset into two subsets, “Tor Traffic” and “NonTor Traffic”, and classify 8 distinct application services within each. This task tests the model’s efficacy in captu...

    2025

  64. [67]

    Tor enforces fixed- size cell padding (512 bytes) and multi-hop randomized delays (jitter)

    Erosion of Statistical Fingerprints.Traditional identifica- tion of streaming media (Audio) and short-burst messaging (Email) Table 5: Mapping between fine-grained categories and merged classes in the CICIoMT2024 and CICIoT2023 datasets.CICIoMT2024 Merged Categories Fine-grained Categories Benign BenignTraffic DDoS MQTT-DDoS-Connect_Flood, MQTT-DDoS-Publi...

  65. [68]

    Protocol Gram- mar

    Low Information Density in Input Window.Our model utilizes the first𝐾= 10packets of a flow. For complex interactive protocols (VoIP/P2P), this window captures rich “Protocol Gram- mar”, such as authentication handshakes and signaling exchanges. However, Audio and Email protocols often exhibit low interaction density during the setup phase. In the first 10...

  66. [69]

    Many-to-One

    A “Many-to-One” Mapping Problem.From an information- theoretic perspective, Tor’s obfuscation creates a “Many-to-One” mapping where distinct application-layer behaviors (continuous streaming vs. bulk downloading vs. message sending) are mapped to an identical physical-layer representation (a sequence of fixed-size encrypted blocks with randomized timing)....

    arXiv 2025

  67. [70]

    divide-and-conquer

    Mitigation of Intra-class Distribution Shift (Proportion- shift):Traffic-MoEcontinues to outperform all competitors, sug- gesting that the MoE architecture effectively mitigates thegradient starvationof minority classes typically encountered in dense net- works. The “divide-and-conquer” dynamic routing allows long-tail minor components to be handled by sp...

  68. [71]

    flooding

    Adaptation to Unseen Variants (Compose-shift):Traffic- MoEranks first in core metrics on both datasets. Notably, while FS- Net achieved higher Macro-Precision inCICIoMT2024,Traffic-MoE surpasses it by a significant margin of14.19%in Macro-Recall. High recall on unseen sub-classes indicates thatTraffic-MoEhas moved beyond simplesignature memorizationto ach...

  69. [2018]

    In2018 IEEE European symposium on security and privacy (EuroS&P)

    Sok: Security and privacy in machine learning. In2018 IEEE European symposium on security and privacy (EuroS&P). IEEE, 399–414

  70. [2022]

    InProceedings of the ACM Web Conference

    ET-BERT: A contextualized datagram representation with pre-training transformers for encrypted traffic classification. InProceedings of the ACM Web Conference. 633–642

  71. [2025]

    In2025 IEEE Symposium on Security and Privacy (SP)

    Trafficformer: an efficient pre-trained model for traffic data. In2025 IEEE Symposium on Security and Privacy (SP). IEEE, 1844–1860

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.