Pith. sign in

REVIEW 4 major objections 4 minor 55 references

The Sweet Danger of Sugar: Debunking Representation Learning for Encrypted Traffic Classification

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Data leakage, not learned representations, drives near-perfect encrypted-traffic classification results.

desk verdict The per-packet split leak is real and well demonstrated, but the paper's sweeping 'representations are uninformative' claim goes beyond its own Table 9. read the letter →

arxiv 2507.16438 v1 pith:QHQFPNE6 submitted 2025-07-22 cs.NI cs.LG

Yuqi Zhao , Giovanni Dettori , Matteo Boffa , Luca Vassio , Marco Mellia This is my paper
classification cs.NIcs.LG
keywords encryptedtrafficclassificationrepresentationlearningshortcutdataleakagetrain-testsplitper-flowevaluationTLSPcap-Encoder
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 argues that the near-perfect accuracies promised by representation-learning models for encrypted traffic classification are largely artifacts of flawed dataset preparation. It shows that splitting packets randomly instead of by connection lets models exploit hidden identifiers - TCP sequence numbers, acknowledgements, and timestamps - that tie test packets to training flows, creating shortcuts that vanish in real deployments. Under a per-flow split with frozen encoders, published models drop to below 40% accuracy on the 120-website TLS task, and randomly re-initializing one model's weights performs as well as its pre-trained weights. The paper offers a header-only T5-based encoder, Pcap-Encoder, as the only tested model whose frozen representation retains signal, but plain machine-learning baselines on hand-selected header fields match or beat it. The upshot is a call for per-flow splitting, frozen-encoder tests, and shallow baselines as the standard evaluation design.

What carries the argument

The mechanism that carries the argument is the implicit flow identifier: TCP sequence and acknowledgement numbers initialize to a random 32-bit space per connection, and TCP timestamp options are close within a session, so every packet in a flow carries fingerprints that tie it to that flow. Under per-packet splits, the classifier can memorize these fingerprints and map a test packet to a training flow of the same class. The paper's corrective machinery is the per-flow split, which confines each flow's packets to one partition, combined with freezing the pre-trained encoder during downstream training so that performance reflects the representation itself rather than task-specific re-training. Pcap-Encoder, the paper's counterproposal, is a T5-based sequence-to-sequence model pre-trained in two self-supervised phases - packet autoencoding and eight header question-answering tasks - to extract semantics from protocol headers only.

What would settle it

Re-run the TLS-120 benchmark with ET-BERT's released pre-trained weights and fine-tuning recipe, but split the data per flow and freeze the encoder; the paper predicts accuracy below 40% (it measures 10.9%). If an independent run under those conditions exceeds 50% accuracy on a held-out set with no flow overlap, the shortcut-leakage explanation would be falsified. A sharper check: keep the original per-packet split but shuffle the TCP sequence numbers, acknowledgement numbers, and timestamp options within each flow before training and testing; under the paper's story, accuracy should collapse from roughly 97% toward 20%, and if it does not, the claimed leakage mechanism is wrong.

Watch

Extended reading notes

Core claim

The central discovery is that data preparation, not model architecture, explains the spectacular results in encrypted traffic classification. With a per-packet split, a packet and other packets from the same TCP flow can land on both sides of the train/test boundary; the model then learns to read implicit flow IDs - closely ranged sequence and acknowledgement number pairs and TCP timestamps - and, because flow membership determines the class label, this leaks the answer. Repeating the same evaluation with a per-flow split (all packets of a flow in one partition) and keeping the pre-trained encoder frozen collapses accuracy: on TLS-120, ET-BERT goes from 97.4% to 10.9%, YaTC to 15.5%, NetMamba to 8.8%, TrafficFormer to 29.7%, and netFound to 1.9%. Randomizing SeqNo, AckNo, and TCP timestamps drops the unfrozen per-packet ET-BERT from 97.4% to 19.5%, and replacing pre-trained weights with random ones leaves accuracy essentially unchanged at 97.1%. Pcap-Encoder, built on T5 and pre-trained to answer questions about packet headers while ignoring encrypted payloads, retains 71.0% accuracy on TLS-120 with a frozen encoder, but Random Forest and gradient-boosted trees fed handpicked header fields score higher still.

Load-bearing premise

The benchmark assumes that a per-flow split with permanently frozen encoders is the correct way to measure representation quality, and that forcing flow-based models to accept a repeated single packet as an artificial flow is a fair comparison; if real deployments allow retraining on deployment traffic, or if per-session or per-client splits change what information leaks, the conclusion that pre-trained representations are uninformative could weaken.

Editorial extensions

If this is right

  • The published over-90% accuracies for ET-BERT, YaTC, NetMamba, TrafficFormer, and netFound on TLS-120 arise only under per-packet splits with unfrozen fine-tuning; under per-flow splits with frozen encoders, accuracy falls to between 1.9% and 29.7%.
  • Removing implicit flow IDs from the data removes the main shortcut: ET-BERT's TLS-120 accuracy drops from 97.4% to 19.5% when sequence numbers, acknowledgement numbers, and timestamps are randomized, confirming that the model was reading flow identity rather than traffic content.
  • Initializing ET-BERT's weights randomly performs on par with its pre-trained weights (97.1% versus 97.4% accuracy) in the per-packet unfrozen setup, indicating that the pre-training phase contributes little to the downstream result.
  • Pcap-Encoder, which reads only protocol headers and ignores encrypted payloads, is the only tested model whose frozen representation supports classification, yet its complexity and training time are much higher than shallow baselines that achieve comparable or better scores.
  • Correct evaluation should freeze encoders, split data per flow, avoid class-support and dataset-reuse filters, and compare against simple baselines; otherwise shortcut learning will continue to inflate reported performance.

Reading between the lines

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

  • Beyond the paper, the per-flow versus per-packet distinction is a general blocking problem: any element-level split of clustered data (sessions, users, locations) can leak cluster identity through incidental identifiers, and the paper's ablation gives a template for diagnosing such leakage in other sequence-classification benchmarks.
  • Beyond the paper, the header-only success of Pcap-Encoder predicts that for fully encrypted traffic, useful representations will come from modelling the plaintext protocol envelope (packet sizes, timing, TLS handshake fields) rather than payload content; a direct testable extension is whether payload-masked pre-training matches or exceeds payload-visible pre-training on TLS-120.
  • Beyond the paper, the frozen-encoder protocol could be adopted as a cheap gate for future representation-learning claims: if a model's embeddings cannot support a shallow classification head without weight updates, the pre-training phase is not contributing. The paper recommends this practice, but making it a universal reporting requirement is an editorial step beyond its claims.
  • The paper leaves open the possibility that a deployed system might fine-tune on deployment traffic, in which case frozen-encoder scores understate usable real-world performance; what the results definitively undermine is the claim that pre-training itself produces the accuracy.
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.

Referee Report

4 major / 4 minor

Summary. The paper challenges the reported high accuracy of representation learning models for encrypted traffic classification (e.g., ET-BERT, YaTC, NetMamba, TrafficFormer, netFound). The authors argue that per-packet dataset splits allow models to exploit implicit flow identifiers (SeqNo, AckNo, TCP timestamps) as shortcut features, and that fine-tuning with unfrozen encoders effectively re-trains the models from scratch. Under their proposed per-flow split with frozen encoders, accuracy on the TLS-120 packet task collapses to below 40%, while removing the implicit flow identifiers from ET-BERT drops its per-packet-split accuracy by roughly 80 percentage points. The paper also introduces Pcap-Encoder, a T5-based encoder trained on protocol headers, which performs better than prior representation learning models under the frozen per-flow protocol, but is matched or exceeded by shallow feature-based baselines. The authors release code, datasets, and models.

Significance. If the claims are appropriately qualified, this is a valuable reproducibility and benchmarking study for the network-traffic-classification community. Strengths include the public release of code, datasets, and checkpoints; the use of several independent tasks and datasets; careful ablations for ET-BERT (Table 6) and Pcap-Encoder (Table 7); feature-importance analysis for the shallow baseline (Fig. 5); and a concrete, reproducible evaluation pipeline with k-fold cross-validation. The paper would have high practical impact if it simply demonstrated that per-packet splitting causes leakage and that reported accuracies are therefore inflated. The more sweeping claim that all pre-trained traffic representations are uninformative, however, is not fully supported by the experiments as reported.

major comments (4)
  1. [§5 (Downstream models)] The packet-level evaluation of flow-based encoders uses an artificial input: for YaTC, NetMamba, TrafficFormer, and netFound, the same single packet is repeated five times to form a synthetic flow, and for netFound the maximum input is filled with the same packet. These models were pre-trained on real flow structure, packet directions, inter-arrival times, and flow lengths, so the repeated-packet input is far outside their pre-training distribution. The large accuracy drops in Tables 3-5 for these models may therefore be caused by input-protocol mismatch rather than by uninformative representations. The authors should either evaluate flow-based encoders on natural flow contexts even for the packet-level task, or explicitly qualify the conclusion to "with repeated-packet input" and justify why this protocol is a fair test of representation quality.
  2. [Abstract/§6.2, Table 9] The abstract and conclusion claim that without shortcuts "these models perform poorly" and that Pcap-Encoder is "the only model that provides an instrumental representation." This is contradicted by Table 9, where netFound reaches 90.8 macro-F1 on TLS-120 flow classification with an unfrozen encoder under the per-flow split, and several other models reach 74-78 macro-F1. The universal statement should be restricted to the packet-level, frozen-encoder, repeated-packet setting, or the flow-level unfrozen results need to be reconciled with the claim. A concrete test would be to report frozen-encoder flow-level results for all models using their natural input format, which would isolate representation quality from fine-tuning effects.
  3. [§4.1 (Dataset Splitting)] The paper correctly notes that "more advanced splits are possible: per-session, per-client, per-location, per-time split, etc.," but then limits the benchmark to per-packet and per-flow splits. The abstract's phrase "as in real scenarios" is therefore an assertion, not a demonstrated result: a per-flow split removes intra-flow leakage but does not address cross-session or cross-client correlations that may also constitute shortcuts in deployment. The authors should either run a per-session or per-client split as an additional stress test, or soften the real-scenario language in the abstract and conclusions.
  4. [§4.2 and Tables 4-5] The paper treats frozen-encoder accuracy as the definition of whether a representation is meaningful, and uses this to conclude that pre-training is "mostly useless." Frozen evaluation is a legitimate and useful stress test, but end-to-end fine-tuning is the standard way these pretrained encoders are used in practice. The evidence that fine-tuning "destroys" pretrained information is indirect: the w/o Pre-training row of Table 6 is measured under a per-packet split where leakage dominates, so it does not isolate the value of pre-training in the realistic per-flow setting. Reporting a linear probe or k-NN evaluation on frozen embeddings, and preferably a per-flow w/o Pre-training control, would make the claim about representation quality more robust. As it stands, the strong conclusion should be qualified to "frozen linear/MLP probing fails" rather than "the representation is uninformative."
minor comments (4)
  1. [Table 3 caption] The caption says "three SoA models" but the table reports results for six models; the caption should be corrected.
  2. [§4.1 vs §6.2] The paper argues against minimum-size filters, but Section 6.2 states that flow-level experiments keep only flows with at least 5 packets. This tension should be clarified by explaining that the flow-length cutoff is a task-definition choice for flow encoders rather than a data-cleaning filter.
  3. [§3.2] The text says ET-BERT uses ISCX-VPN and "(likely) CSTNET-TLS1.3" for both upstream and downstream training; since the authors inspected the original code and models, the uncertain "likely" should be resolved or removed.
  4. [Appendix A.1.3] The Q&A pre-training reports 98.2% average accuracy on the question-answering test set, but no split or leakage-prevention details are given for this internal test; a sentence describing the split would improve confidence in the number.

Circularity Check

0 steps flagged · score 0.0 of 10

The debunking is empirical and self-contained: performance collapses are measured on open datasets with public checkpoints, and the leakage diagnosis is an ablation, not an identity.

full rationale

No circular step meets the evidentiary bar. The paper's central claims are empirical findings: per-flow splitting with frozen encoders yields low accuracy (Table 3), per-packet splitting with unfrozen encoders yields high accuracy (Table 5), and removing SeqNo/AckNo/timestamps in the test set drops ET-BERT from 97.4% to 19.5% accuracy (Table 6). These are controlled experiments, not quantities defined in terms of each other. The frozen-encoder protocol is proposed as a methodology, not derived from the models under test, and it is supported by independent evidence such as the 5-NN purity analysis (Figure 4) and the random-initialization ablation (Table 6, 'w/o Pre-training'). Pcap-Encoder's 'instrumental' status is an empirical outcome, and the paper's own ablation (Table 11) shows that removing its pre-training components degrades performance, so the claim is not true by construction. The only self-citation is reference [6], a technical report for Pcap-Encoder implementation details, which is not load-bearing for the central negative result. The main limitations are threats to external validity rather than circularity: the paper equates per-flow splitting with 'real scenarios' while acknowledging per-session, per-client, and other splits as possible but not evaluating them, and it feeds flow-oriented encoders a repeated single packet for packet classification, which is outside their pre-training distribution. These choices could bias the magnitude of the reported collapse, but they do not make the derivation equivalent to its inputs. The paper is self-contained against external benchmarks and public checkpoints, so the appropriate score is 0.

Assumptions & free parameters 4 free parameters · 7 assumptions · 0 invented entities

The paper carries no hidden fitted constants in the usual sense; its conclusions rest on evaluation-protocol axioms instead. The most consequential choices are the per-flow/frozen regime, the undersampled training distribution, and the hand-crafted baseline feature set, each of which is transparently described and partially controlled by ablation experiments.

free parameters (4)
  • Frozen-encoder hyperparameters for prior models = ET-BERT 2e-3/60; YaTC/NetMamba 2e-3/200; TrafficFormer 1e-4/60; netFound 2.5e-6/100
    Original papers did not define frozen-encoder training, so these learning rates and epochs are hand-chosen by the authors; they directly affect the claim that frozen representations fail.
  • Hand-crafted shallow-model feature set = IPv4/IPv6 and UDP/TCP header fields, incl. IP addresses, TCP Seq/Ack, timestamps (Table 12)
    The baselines that outperform Pcap-Encoder use expert-selected features; the w/o IP address column partly controls for this, but the feature set itself is a design choice.
  • Training undersampling to minority class = Each class reduced to the size of the minority class (Section 5)
    This stress-test changes the training distribution from the original papers and lowers performance for all models; the authors note an unbalanced 8:1:1 split increases scores.
  • Flow-level minimum length = 5 packets (Section 6.2)
    Flows with fewer than five packets are excluded, which the authors call the best case for encoders but is still an arbitrary threshold that changes the task.
assumptions (7)
  • domain assumption Robust encryption implies no learnable correlation among encrypted payload bytes
    Invoked in Section 3.3 to argue that masked-autoencoder pre-training on payloads makes little sense and to justify Pcap-Encoder ignoring payloads.
  • domain assumption Per-flow split with all packets of a flow kept together simulates real deployment and removes implicit flow-ID shortcuts
    This is the central evaluation premise introduced in Section 4.1; if per-session or per-client leakage is the real risk, the benchmark conclusions could shift.
  • domain assumption Freezing the encoder is the appropriate way to measure whether pre-training learned a useful representation
    Section 4.2 advocates frozen encoders; the claim that previous representations are uninformative depends on this definition.
  • ad hoc to paper Repeating a single packet five times produces a valid flow input for flow-based encoders
    Section 5 applies this to ET-BERT, YaTC, NetMamba, and other models in packet-level tasks; it may handicap models whose pretraining expects true flow context.
  • domain assumption Trace-level class labels apply to every packet and flow in the trace
    Used in Section 4.2 with the footnote 'Even if questionable, this is the same formulation previous works used'; noisy labels would add label noise to all models equally.
  • domain assumption Public CSTN-TLS1.3 dataset contains no plaintext SNI
    Section 4.1 footnote 7: the original ET-BERT paper says SNI is present but the public dataset is not; this affects whether TLS-120 is truly an everything-encrypted task.
  • domain assumption The five selected models (ET-BERT, YaTC, NetMamba, TrafficFormer, netFound) are representative of the representation learning approaches in this area
    Section 5 selects them as representative; the overclaim that Pcap-Encoder is the only instrumental model depends on this coverage.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Sweet Danger of Sugar: Debunking Representation Learning for Encrypted Traffic Classification." pith.science (2026). https://pith.science/paper/QHQFPNE6

@misc{pith2026250716438,
  author       = {Pith},
  title        = {Pith review of: The Sweet Danger of Sugar: Debunking Representation Learning for Encrypted Traffic Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QHQFPNE6}},
  note         = {Machine review of arXiv:2507.16438}
}
read the original abstract

Recently we have witnessed the explosion of proposals that, inspired by Language Models like BERT, exploit Representation Learning models to create traffic representations. All of them promise astonishing performance in encrypted traffic classification (up to 98% accuracy). In this paper, with a networking expert mindset, we critically reassess their performance. Through extensive analysis, we demonstrate that the reported successes are heavily influenced by data preparation problems, which allow these models to find easy shortcuts - spurious correlation between features and labels - during fine-tuning that unrealistically boost their performance. When such shortcuts are not present - as in real scenarios - these models perform poorly. We also introduce Pcap-Encoder, an LM-based representation learning model that we specifically design to extract features from protocol headers. Pcap-Encoder appears to be the only model that provides an instrumental representation for traffic classification. Yet, its complexity questions its applicability in practical settings. Our findings reveal flaws in dataset preparation and model training, calling for a better and more conscious test design. We propose a correct evaluation methodology and stress the need for rigorous benchmarking.

Figures

Figures reproduced from arXiv: 2507.16438 by the authors.

Figure 1
Figure 1. Accuracy of classifiers evaluated (TLS-120 dataset, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Schema of Pcap-Encoder, our proposal. Phase 1: Encoder update. We first update the original T5 en￾coder using raw packet traces to adapt it to the new data format and semantics. The encoder’s goal is to map the original data into a numerical space, possibly removing redundant or useless (e.g., constant) information. For this, we train the (base) T5 to reconstruct the original packet from the internal representation.… view at source ↗
Figure 3
Figure 3. Per-flow and per-packet split. ports [27, 33, 42, 49, 53, 54]. For example, YaTC randomises the IP address and sets the port number to zero; PacRep and NetMamba set both IP address and port to zero; PTU removes IP address, MAC address and checksum; TrafficFormer randomises IP address and ports, and some specific fields (such as timestamp) for data augmen￾tation. ET-BERT removes the IP header entirely. netFound omits… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: 5-NN purity of embeddings for ET-BERT. With a frozen encoder, 71% of points do not have a sample of the same class as TOP-5 neighbour. Situation changes only when the encoder is unfrozen. Scenario Dataset AC F1 Original 97.4 96.8 w/o SeqNo/AckNo w/o Timestamp (only tes…
Figure 5
Figure 5. Figure 5: Feature importance for the Random Forest model [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Relative training and inference times. All models [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

55 extracted references · 43 canonical work pages

  1. [1]

    Daniel Arp, Erwin Quiring, Feargus Pendlebury, Alexander Warnecke, Fabio Pierazzi, Christian Wressnegger, Lorenzo Cavallaro, and Konrad Rieck. 2022. Dos and don’ts of machine learning in computer security. In 31st USENIX Security Symposium (USENIX Security 22) . 3971–3988

    2022

  2. [2]

    Kenjiro Cho, Koushirou Mitsuya, and Akira Kato. 2000. Traffic data repository at the{WIDE} project. In 2000 USENIX Annual Technical Conference (USENIX ATC 00)

    work page 2000

  3. [3]

    Noam Chomsky. 1957. Syntactic Structures. De Gruyter Mouton, Berlin, Boston

    work page 1957

  4. [4]

    Alberto Dainotti, Antonio Pescape, and Kimberly C Claffy. 2012. Issues and future directions in traffic classification. IEEE network 26, 1 (2012), 35–40

    work page 2012

  5. [5]

    Ferdinand De Saussure. 2004. Course in general linguistics. Literary theory: An anthology 2 (2004), 59–71

    work page 2004

  6. [6]

    Giovanni Dettori. 2024. Designing and engineering a Q&A LLM for network packet representation. MSc thesis. Politecnico di Torino. Available at https: //webthesis.biblio.polito.it/33158/

    work page 2024

  7. [7]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)

    arXiv 2018

  8. [8]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al . 2021. An Image is Worth 16x16 Words: Trans- formers for Image Recognition at Scale. In International Conference on Learning Representations

    work page 2021

Show all 55 references
  1. [9]

    Ghorbani

    Gerard Draper-Gil, Arash Habibi Lashkari, Mohammad Saiful Islam Mamun, and Ali A. Ghorbani. 2016. Characterization of Encrypted and VPN Traffic using Time-related Features. In International Conference on Information Systems Security and Privacy

    2016

  2. [10]

    Nick Erickson, Jonas Mueller, Alexander Shirkov, Hang Zhang, Pedro Larroy, Mu Li, and Alexander Smola. 2020. AutoGluon-Tabular: Robust and Accurate AutoML for Structured Data. arXiv preprint arXiv:2003.06505 (2020)

    2020 arXiv

  3. [11]

    Amir Feder, Katherine A Keith, Emaad Manzoor, Reid Pryzant, Dhanya Sridhar, Zach Wood-Doughty, Jacob Eisenstein, Justin Grimmer, Roi Reichart, Margaret E Roberts, et al. 2022. Causal inference in natural language processing: Estima- tion, prediction, interpretation and beyond....

    2022

  4. [12]

    Michael Finsterbusch, Chris Richter, Eduardo Rocha, Jean-Alexander Muller, and Klaus Hanssgen. 2013. A survey of payload-based traffic classification approaches. IEEE Communications Surveys & Tutorials 16, 2 (2013), 1135–1156

    2013

  5. [13]

    Robert Flood, Gints Engelen, David Aspinall, and Lieven Desmet. 2024. Bad design smells in benchmark nids datasets. In 2024 IEEE 9th European Symposium on Security and Privacy (EuroS&P) . IEEE, 658–675

    2024

  6. [14]

    George Forman and Martin Scholz. 2010. Apples-to-apples in cross-validation studies: pitfalls in classifier performance measurement. Acm Sigkdd Explorations Newsletter 12, 1 (2010), 49–57

    2010

  7. [15]

    Robert Geirhos, Jörn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. 2020. Shortcut learning in deep neural networks. Nature Machine Intelligence 2, 11 (2020), 665–673

    2020

  8. [16]

    Albert Gu and Tri Dao. 2023. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752 (2023)

    2023 arXiv

  9. [17]

    Satyandra Guthula, Roman Beltiukov, Navya Battula, Wenbo Guo, and Arpit Gupta. 2023. netFound: Foundation model for network security. arXiv preprint arXiv:2310.17025 (2023)

    2023 arXiv

  10. [18]

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

    2020

  11. [19]

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Gir- shick. 2022. Masked autoencoders are scalable vision learners. In 2022 IEEE/CVF conference on computer vision and pattern recognition . 16000–16009

    2022

  12. [20]

    Jeremy Howard and Sebastian Ruder. 2018. Universal language model fine-tuning for text classification. arXiv preprint arXiv:1801.06146 (2018)

    2018 arXiv

  13. [21]

    Arthur S Jacobs, Roman Beltiukov, Walter Willinger, Ronaldo A Ferreira, Arpit Gupta, and Lisandro Z Granville. 2022. Ai/ml for network security: The emperor has no clothes. In2022 ACM SIGSAC Conference on Computer and Communications Security. 1537–1551

    2022

  14. [22]

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. 2020. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361 (2020)

    2020 arXiv

  15. [23]

    Will Knight. 2023. Google’s Gemini Is The Real Start of the Generative AI Boom. https://www.wired.com/story/google-gemini-generative-ai-boom/

    2023

  16. [24]

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. 2012. Imagenet classifi- cation with deep convolutional neural networks. Advances in neural information processing systems 25 (2012)

    2012

  17. [25]

    Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2019. Albert: A lite bert for self-supervised learning of language representations. arXiv preprint arXiv:1909.11942 (2019)

    2019 arXiv

  18. [26]

    Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. 2022. Scaling & shifting your features: A new baseline for efficient model tuning. Advances in Neural Information Processing Systems 35 (2022), 109–123

    2022

  19. [27]

    Xinjie Lin, Gang Xiong, Gaopeng Gou, Zhen Li, Junzheng Shi, and Jing Yu. 2022. Et-bert: A contextualized datagram representation with pre-training transform- ers for encrypted traffic classification. In 2022 ACM Web Conference. 633–642. Debunking Representation Learning

    2022

  20. [28]

    Shayne Longpre, Le Hou, Tu Vu, Albert Webson, Hyung Won Chung, Yi Tay, Denny Zhou, Quoc V Le, Barret Zoph, Jason Wei, et al. 2023. The flan collection: Designing data and methods for effective instruction tuning. In International Conference on Machine Learning . PMLR, 22631–22648

    2023

  21. [29]

    Minh-Thang Luong, Hieu Pham, and Christopher D Manning. 2015. Effec- tive approaches to attention-based neural machine translation. arXiv preprint arXiv:1508.04025 (2015)

    2015 arXiv

  22. [30]

    Ben Mann, Nick Ryder, Melanie Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems 33 (2020), 1877–1901

    2020

  23. [31]

    David Marr. 2010. Vision: A computational investigation into the human represen- tation and processing of visual information . MIT press

    2010

  24. [32]

    Thomas McCoy, Ellie Pavlick, and Tal Linzen

    R. Thomas McCoy, Ellie Pavlick, and Tal Linzen. 2019. Right for the Wrong Reasons: Diagnosing Syntactic Heuristics in Natural Language Inference. In Annual Meeting of the Association for Computational Linguistics

    2019

  25. [33]

    Xuying Meng, Yequan Wang, Runxin Ma, Haitong Luo, Xiang Li, and Yujun Zhang. 2022. Packet representation learning for traffic classification. In 2022 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 3546–3554

    2022

  26. [34]

    Sam Meredith. 2023. A ‘thirsty’ generative AI boom poses a growing problem for Big Tech. https://www.cnbc.com/2023/12/06/water-why-a-thirsty-generative- ai-boom-poses-a-problem-for-big-tech.html

    2023

  27. [35]

    Jose G Moreno-Torres, Troy Raeder, Rocío Alaiz-Rodríguez, Nitesh V Chawla, and Francisco Herrera. 2012. A unifying view on dataset shift in classification. Pattern recognition 45, 1 (2012), 521–530

    2012

  28. [36]

    Cherry picking

    Janice M Morse. 2010. “Cherry picking”: Writing from thin data. 3–3 pages

    2010

  29. [37]

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

    2015

  30. [38]

    David Naylor, Alessandro Finamore, Ilias Leontiadis, Yan Grunenberger, Marco Mellia, Maurizio Munafò, Konstantina Papagiannaki, and Peter Steenkiste. 2014. The cost of the" s" in https. In 2014 10th ACM International on Conference on Emerging Networking Experiments and Technol...

    2014

  31. [39]

    Thuy TT Nguyen and Grenville Armitage. 2008. A survey of techniques for internet traffic classification using machine learning. IEEE communications surveys & tutorials 10, 4 (2008), 56–76

    2008

  32. [40]

    Fannia Pacheco, Ernesto Exposito, Mathieu Gineste, Cedric Baudoin, and Jose Aguilar. 2018. Towards the deployment of machine learning solutions in network traffic classification: A systematic survey. IEEE Communications Surveys & Tutorials 21, 2 (2018), 1988–2014

    2018

  33. [41]

    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

  34. [42]

    Lingfeng Peng, Xiaohui Xie, Sijiang Huang, Ziyi Wang, and Yong Cui. 2024. PTU: Pre-trained Model for Network Traffic Understanding. In 2024 32nd IEEE International Conference on Network Protocols (ICNP)

    2024

  35. [43]

    Jing Qin. 2017. Biased sampling, over-identified parameter problems and beyond . Vol. 5. Springer

    2017

  36. [44]

    Joaquin Quiñonero-Candela, Masashi Sugiyama, Anton Schwaighofer, and Neil D Lawrence. 2022. Dataset shift in machine learning . Mit Press

    2022

  37. [45]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research 21, 140 (2020), 1–67

    2020

  38. [46]

    Shahbaz Rezaei and Xin Liu. 2019. Deep learning for encrypted traffic classifica- tion: An overview. IEEE communications magazine 57, 5 (2019), 76–81

    2019

  39. [47]

    Tatiana Tommasi, Novi Patricia, Barbara Caputo, and Tinne Tuytelaars. 2017. A deeper look at dataset bias. Domain adaptation in computer vision applications (2017), 37–55

    2017

  40. [48]

    Martino Trevisan, Danilo Giordano, Idilio Drago, Marco Mellia, and Maurizio Munafo. 2020. Five Years at the Edge: Watching Internet From the ISP Network. IEEE/ACM Transactions on Networking 28, 02 (2020), 561–574

    2020

  41. [49]

    Tongze Wang, Xiaohui Xie, Wenduo Wang, Chuyi Wang, Youjian Zhao, and Yong Cui. 2024. NetMamba: Efficient Network Traffic Classification via Pre-training Unidirectional Mamba. In 2024 32nd IEEE International Conference on Network Protocols (ICNP)

    2024

  42. [50]

    Wei Wang, Ming Zhu, Xuewen Zeng, Xiaozhou Ye, and Yiqiang Sheng. 2017. Mal- ware traffic classification using convolutional neural network for representation learning. In 2017 International Conference on Information Networking (ICOIN) . 712–717

    2017

  43. [51]

    Nimesha Wickramasinghe, Arash Shaghaghi, Gene Tsudik, and Sanjay Jha. 2025. SoK: Decoding the Enigma of Encrypted Network Traffic Classifiers. In 2025 IEEE Symposium on Security and Privacy (SP) . IEEE, 1825–1843

    2025

  44. [52]

    Walter Willinger, Ronaldo A Ferreira, Arpit Gupta, Roman Beltiukov, Satyandra Guthula, Lisandro Z Granville, and Arthur S Jacobs. 2025. When Something Looks Too Good To Be True, It Usually Is! AI Is Causing A Credibility Crisis In Networking. ACM SIGCOMM Computer Communication...

    2025

  45. [53]

    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. In2023 37th AAAI Conference on Artificial Intelligence, Vol. 37. ...

    2023

  46. [54]

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

  47. [2024]

    In 2025 IEEE Symposium on Security and Privacy (SP)

    TrafficFormer: An Efficient Pre-trained Model for Traffic Data. In 2025 IEEE Symposium on Security and Privacy (SP) . IEEE Computer Society, 102–102. A Appendix Appendices are supporting material that has not been peer-reviewed. A.1 Pcap-Encoder details A.1.1 Packet representa...

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.