Energy-Efficient Deep Learning for Traffic Classification on Microcontrollers

Adel Chehade; Edoardo Ragusa; Paolo Gastaldo; Rodolfo Zunino

REVIEW 2 major objections 6 minor 31 references

Energy-Efficient Deep Learning for Traffic Classification on Microcontrollers

T0 review · 2 major / 6 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 A hardware-aware search produces a microcontroller-compatible 1D-CNN that reports 96.59% accuracy on the ISCX VPN-nonVPN task.

desk verdict Real measured deployment on STM32 boards, but the headline accuracy may be a validation-set number due to no independent test split. read the letter →

arxiv 2506.10851 v1 pith:KEZOQFY2 submitted 2025-06-12 cs.NI cs.LG

Adel Chehade , Edoardo Ragusa , Paolo Gastaldo , Rodolfo Zunino This is my paper · ORCID

classification cs.NIcs.LG

keywords trafficclassificationmicrocontrollershardware-awareneuralarchitecturesearch1D-CNNINT8quantizationenergy-efficientinferenceencryptedIoTsecurity

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

Encrypted traffic classification usually requires either payload inspection or heavy cloud-side deep models. This paper tries to establish that a small neural network can do the job on a microcontroller, where memory, compute, and energy are tightly limited. The route is hardware-aware neural architecture search over 1D convolutional networks, constrained by Flash, RAM, and FLOPs budgets. The search returns a model with 88.26K parameters and 10.08M FLOPs that reaches 96.59% accuracy on the ISCX VPN-nonVPN session-level task and 94–99% on four related tasks; after 8-bit quantization it runs in 31.43 ms at 7.86 mJ per inference on a high-performance board and 115.40 ms at 29.10 mJ on a budget board. If these numbers hold on genuinely unseen test data, the immediate consequence is practical on-device encrypted traffic analysis for low-power IoT security.

What carries the argument

The machinery is an evolutionary hardware-aware architecture search over modular 1D-CNN blocks, where each candidate must satisfy explicit budgets: parameter count below a Flash limit, maximum tensor size below a RAM limit, and FLOPs below a compute limit. The search starts from a parent architecture and creates mutated children by adding or removing blocks and changing filters, kernel sizes, strides, pooling, dropout, and padding; invalid candidates are discarded without training. One hundred generations of ten candidates each are trained and scored on a 20% validation split, and the best child becomes the next parent. The selected architecture has three convolutional layers with 129, 110, and 38 filters, mixed pooling, global average pooling, and a dense softmax. Deployment uses 8-bit integer quantization and automatic C code generation for the target boards, with latency and energy measured by running 1000 inferences and recording current draw above idle.

What would settle it

Rerun the final architecture on a held-out split of ISCX VPN-nonVPN that was never used during the 100-generation search, and compare against 96.59%. If accuracy drops materially, the headline result reflected selection on the validation set; if it holds, the central claim survives.

Watch

Extended reading notes

Core claim

The central claim is that a 1D-CNN produced by hardware-aware evolutionary search can classify session-level encrypted traffic at near-state-of-the-art accuracy with a footprint small enough for microcontrollers. On the primary VPN-nonVPN task the model reports 96.59% accuracy and 96.54% F1 with 88.26K parameters, a 20.12K maximum tensor size, and 10.08M FLOPs. The same architecture, retrained, reports 99.95% on VPN protocol detection, 99.19% on VPN traffic type, 94.17% on non-VPN traffic type, and 96.94% on network usage categorization. Integer 8-bit quantization cuts 1–2 accuracy points, and the deployed model is measured at 31.43 ms and 7.86 mJ per inference on the STM32F746G-DISCO and 115.40 ms and 29.10 mJ on the Nucleo-F401RE. The paper's point is that this combination—accuracy, generalization, and energy cost—makes on-device encrypted traffic analysis feasible.

Load-bearing premise

The argument stands on the assumption that the reported accuracies come from a test set that played no role in the architecture search. The paper describes only a 20% validation holdout used to select the best model, so if the reported numbers were computed on that same holdout, true generalization would be lower than stated.

Editorial extensions

If this is right

Session-level encrypted traffic analysis can run in 31 ms at 7.86 mJ on a 216 MHz microcontroller and in 115 ms at 29.10 mJ on a lower-end board.
The 88.26K-parameter, 20.12K-tensor footprint fits the 512 KB Flash / 96 KB RAM class, so no cloud round-trip is needed for classification.
INT8 quantization is an acceptable deployment step: 1–2 accuracy points lost across tasks, with accuracy still in the 93–100% range.
One search produces a model that transfers to four other traffic-classification tasks without re-running the search.
Latency and energy figures are low enough for real-time IoT monitoring as long as session aggregation has already happened.

Reading between the lines

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

Because the paper reports only a 20% validation split for architecture selection and no separate final test set, a reasonable next step is a nested cross-validation rerun; the 96.59% figure would be convincing only if it survives on never-seen folds.
The 31–115 ms latencies are well above per-packet wire rates, so the practical deployment target is session-level decision points (firewalls, edge gateways) rather than line-rate inspection; the paper itself notes the aggregation delay of session collection.
The energy measurement subtracts idle current, so the reported mJ figures describe incremental inference cost; total device energy for battery-powered nodes would be higher and might change the board ranking.
An immediate testable extension is applying the same HW-NAS pipeline to other encrypted-traffic datasets (e.g., Tor traffic or malware flows) to see whether the searched 1D-CNN shape transfers beyond ISCX-derived tasks.

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

Colleague,

Worth a look if you care about embedded ML for traffic classification: the authors did a real hardware-aware NAS search, deployed the result on two STM32 boards, and report measured latency and energy. That part is solid and useful. The architecture (three 1D-CNN layers, 88K params, 10M FLOPs) is new relative to their prior work, and the four-task generalization table shows the model transfers reasonably well. The deployment numbers are the real contribution: 31.43 ms / 7.86 mJ on the STM32F746G-DISCO and 115.40 ms / 29.10 mJ on the Nucleo-F401RE are concrete and hard to get from simulation.

The soft spot is the evaluation protocol, and it is load-bearing. Section IV-C states that a 20% holdout validation set is used to guide architecture selection, and Algorithm 1 selects the parent that maximizes accuracy on that validation set. The paper never describes an independent test split. The headline 96.59% on VPN-NonVPN (and the generalization accuracies in Table IV and Figure 1) may therefore all be computed on the same validation set used to pick the architecture after 100 generations times 10 candidates. That repeated selection can inflate accuracy well beyond true generalization. The authors say the selected model is retrained for generalization, but retraining on the same data does not fix selection leakage. The reader's stress-test note is correct: this is not a minor omission.

A few smaller issues: the baseline hardware numbers in Table II are estimates made in Keras from other papers, so the efficiency comparison is not apples-to-apples. The power measurement (USB power meter, idle subtraction) is plausible but coarse; the 0.25 W reading for both boards looks suspiciously round. No code, data, or error bars are provided.

I'd still send this out for review. The deployment measurements and the architecture are genuinely useful to the embedded-ML-for-networking community, and the evaluation flaw is fixable with a proper test split and honest reporting of selection-on-validation effects. A referee should ask for that before acceptance, not reject on principle. The paper is clearly written and the method section is transparent enough that the problem is visible on a careful read, which counts in its favor.

Referee Report

2 major / 6 minor

Summary. The paper proposes a hardware-aware neural architecture search (HW-NAS) approach to design a lightweight 1D-CNN for session-level encrypted traffic classification on microcontrollers. The searched model has 88.26K parameters, a 20.12K maximum tensor size, and 10.08M FLOPs, and is reported to achieve 96.59% accuracy on the ISCX VPN-NonVPN task. The paper also reports generalization across four additional traffic classification tasks, with accuracies between 94% and 99%, and evaluates deployment after INT8 quantization on two microcontroller boards, STM32F746G-DISCO and Nucleo-F401RE, reporting latencies of 31.43 ms and 115.40 ms and energies of 7.86 mJ and 29.10 mJ per inference, respectively.

Significance. If the reported accuracies are confirmed on an independent test split, the paper makes a useful practical contribution: it demonstrates that HW-NAS can produce microcontroller-deployable traffic classifiers with competitive accuracy and very low hardware footprint. The deployment measurements on two real boards, including measured latency and energy, are a concrete strength, as is the evaluation of INT8 quantization. The cross-task generalization tests are also valuable. However, the central accuracy claim is undermined by the absence of a described held-out test set, so the significance of the numerical results currently rests on an unverified evaluation protocol.

major comments (2)

[Section IV-C, Algorithm 1, Eq. (1)] The paper never defines an independent test split. Equation (1) maximizes Accuracyval, Algorithm 1 selects each generation's parent by E(ac, XV), and Section IV-C says a 20% holdout validation set guides architecture selection. Sections V-B and V-C then report the resulting numbers (e.g., 96.59% in Table II and the Float32 column of Table IV) as the model's accuracies, but no untouched test set is described. With about 1000 candidate architectures evaluated against the same validation split during the search, these figures can be inflated by selection-on-validation bias. Please report accuracies on a separate test split, or explicitly state that the reported numbers are validation accuracies and temper the generalization claims accordingly.
[Section IV-C, Table II] The baseline hardware metrics in Table II are described as estimated 'in Keras based on reported architectural details,' but the estimation procedure is not described. The comparison is central to the efficiency claim (e.g., the text states that [28] uses '23 times more parameters and 312 times more FLOPs' than the proposal). Please specify how each baseline's layer configuration, input size, and parameter/tensor/FLOPs counts were obtained, and ideally verify them against the original implementations; otherwise the comparative efficiency numbers are not reproducible and may be artifacts of estimation choices.

minor comments (6)

[Table I] Table I lists 'VPN-NonVPN' with 11 classes, while the text describes the task as distinguishing VPN from Non-VPN traffic; clarify whether this is a binary or multiclass formulation and how the 11 classes are defined.
[Algorithm 1 and Section IV-C] Algorithm 1 returns the final parent ap, but Section IV-C says 'the architecture that achieves the highest validation accuracy is selected'; reconcile the algorithm with the text and describe how the best architecture is retained across generations, or the distinction between 'last parent' and 'best architecture' is ambiguous.
[Eq. (1)] Equation (1) is stated as an optimization with w*(a) = argmin Ltrain(w,a) as a constraint; this is a nested optimization rather than a conventional constraint, and the notation would be clearer if the inner problem were stated separately.
[Section IV-C] The paper states that each candidate architecture is trained three times with a multi-start approach, but it is not reported whether selection uses the mean, the best, or a single run; this choice affects the selection protocol and should be clarified.
[Section VI-B and Table V] Latency measurement is described only as the average over 1000 consecutive inferences; clarify whether the reported values include only the inference call or also any preprocessing and I/O overhead, and explain the small discrepancy between the reported energy and the power-latency product for the Nucleo-F401RE (0.25 W x 115.40 ms = 28.85 mJ versus 29.10 mJ).
[Figure 1] The figure contains NaN entries for some baseline methods in the Traffic-Cat task and an unusual horizontal scale; the caption should explain how missing values were handled and how the accuracy values were obtained.

Circularity Check

1 steps flagged · score 6.0 of 10

The 96.59% headline accuracy is the validation score used to select the architecture; without a separate test set, the generalization claim is partially circular.

fitted input called prediction [Section III-A (Eq. 1), Algorithm 1, Section IV-C, Section V-B]
"The optimization problem seeks to maximize validation accuracy subject to typical embedded-system constraints: max a∈A Accuracyval(w∗(a),a) ... The architecture that achieves the highest validation accuracy is selected as the final model. ... The proposal achieves 96.59% accuracy and an F1 score of 96.54%."

Eq. (1) explicitly sets the search objective to Accuracy_val, and Algorithm 1 makes every parent-selection decision by argmax over E(ac, XV), the 20% validation split. Section IV-C then defines the final model as the architecture with the highest validation accuracy, and Section V-B presents 96.59% as the achieved result. No separate untouched test split is described anywhere; the only 'generalization' step is retraining the selected model on the same tasks. Thus the reported accuracy is the same validation quantity used to choose among about 1,000 candidates (100 generations x 10 children).

full rationale

The only substantive circularity is the accuracy evaluation protocol. Equation (1) defines the search objective as maximizing validation accuracy; Algorithm 1 selects parents and the final architecture by that same validation accuracy; and Section V-B reports 96.59% as the achieved accuracy without describing any separate test split. Thus the headline number is the optimization criterion itself, so the generalization claim is partly reduced to the selection fit. This is a real but localized issue: the deployment measurements (latency, power, energy on two real boards) are independent empirical results, and the cross-task numbers are not used as selection criteria, so they are less affected. The authors' self-citations ([22], [24]-[26]) supply the NAS machinery and thresholds, but the measured results do not reduce to those citations; therefore no separate self-citation circularity is scored. Overall score 6: one central 'prediction' (accuracy/generalization) is not independently evaluated and is the selection objective.

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

The central claim rests on the searched architecture, which is a fitted artifact. Hardware constraint thresholds and search hyperparameters are chosen by hand; baseline resource numbers are estimates; the 1D-CNN search space and the benchmark's label quality are assumed. These are reasonable for an engineering study, but they mean the reported efficiency and accuracy figures are conditional on the search configuration and preprocessing choices, not derived from first principles.

free parameters (4)

NAS-discovered architecture hyperparameters = filters 129/110/38; kernels 7/4/7; strides 5/2/2; pooling choices
Chosen by evolutionary search to maximize validation accuracy (Section V-A); the central model's efficiency and accuracy numbers depend entirely on these.
Hardware constraint thresholds F_Th, R_Th, Flops_Th = not numerically stated in paper
Set by hand from target microcontroller specs and prior work [24], [25] (Section III-A); they shape the final model footprint.
Input session length = 784 bytes
Adopted from [8], [27] for input standardization; a modeling choice that affects accuracy and memory usage.
NAS meta-parameters = Ng=100, Nc=10, filter range 16-140, kernel 3-7, stride 1-6, dropout 0.1-0.5, pool 2-3
Search hyperparameters chosen by hand (Section IV-C); they define the search space and budget, so they influence the outcome.

assumptions (4)

domain assumption 1D-CNNs are effective and efficient for encrypted traffic classification from raw bytes
Invoked in Section III-B based on [8], [9], [27]; if false, the search space already excludes better model families.
domain assumption The evolutionary NAS, with the given generations and candidate counts, finds a near-optimal architecture in the search space
Assumed by Algorithm 1 and the convergence claim in Sections III-C and IV-C; no proof or repeated-search stability analysis is provided.
domain assumption The ISCX VPN-NonVPN dataset provides reliable ground-truth labels and task definitions
Standard benchmark assumption; the paper cites [12] but does not validate label quality or class balance.
domain assumption Keras-based estimates of baseline parameter counts, tensor sizes, and FLOPs are accurate and comparable
Baseline resource usage in Table II is estimated from reported architectural details (Section IV-C), so comparability with the proposed model is assumed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Energy-Efficient Deep Learning for Traffic Classification on Microcontrollers." pith.science (2026). https://pith.science/paper/KEZOQFY2

@misc{pith2026250610851,
  author       = {Pith},
  title        = {Pith review of: Energy-Efficient Deep Learning for Traffic Classification on Microcontrollers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KEZOQFY2}},
  note         = {Machine review of arXiv:2506.10851}
}

read the original abstract

In this paper, we present a practical deep learning (DL) approach for energy-efficient traffic classification (TC) on resource-limited microcontrollers, which are widely used in IoT-based smart systems and communication networks. Our objective is to balance accuracy, computational efficiency, and real-world deployability. To that end, we develop a lightweight 1D-CNN, optimized via hardware-aware neural architecture search (HW-NAS), which achieves 96.59% accuracy on the ISCX VPN-NonVPN dataset with only 88.26K parameters, a 20.12K maximum tensor size, and 10.08M floating-point operations (FLOPs). Moreover, it generalizes across various TC tasks, with accuracies ranging from 94% to 99%. To enable deployment, the model is quantized to INT8, suffering only a marginal 1-2% accuracy drop relative to its Float32 counterpart. We evaluate real-world inference performance on two microcontrollers: the high-performance STM32F746G-DISCO and the cost-sensitive Nucleo-F401RE. The deployed model achieves inference latencies of 31.43ms and 115.40ms, with energy consumption of 7.86 mJ and 29.10 mJ per inference, respectively. These results demonstrate the feasibility of on-device encrypted traffic analysis, paving the way for scalable, low-power IoT security solutions.

Figures

Figures reproduced from arXiv: 2506.10851 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 25 canonical work pages

[28]

Centime: A direct comprehensive traffic features extraction for en- crypted traffic classification,

W. Maonan, Z. Kangfeng, X. Ning, Y . Yanqing, and W. Xiujuan, “Centime: A direct comprehensive traffic features extraction for en- crypted traffic classification,” in Proc. 2021 IEEE 6th Int. Conf. Comput. Commun. Syst. (ICCCS). IEEE, 2021, pp. 490–498

work page 2021
[1]

A novel traffic classification ap- proach by employing deep learning on software-defined networking,

D. Nu ˜nez-Agurto, W. Fuertes, L. Marrone, E. Benavides-Astudillo, C. Coronel-Guerrero, and F. Perez, “A novel traffic classification ap- proach by employing deep learning on software-defined networking,” Future Internet, vol. 16, no. 5, p. 153, 2024

work page 2024
[2]

Glads: A global-local attention data selection model for multimodal multitask encrypted traffic classification of iot,

J. Dai, X. Xu, and F. Xiao, “Glads: A global-local attention data selection model for multimodal multitask encrypted traffic classification of iot,” Computer Networks, vol. 225, p. 109652, 2023

work page 2023
[3]

Cbs: A deep learning approach for encrypted traffic classification with mixed spatio- temporal and statistical features,

M. Seydali, F. Khunjush, B. Akbari, and J. Dogani, “Cbs: A deep learning approach for encrypted traffic classification with mixed spatio- temporal and statistical features,” IEEE Access, 2023

work page 2023
[4]

Deep learning and pre- training technology for encrypted traffic classification: A comprehensive review,

W. Dong, J. Yu, X. Lin, G. Gou, and G. Xiong, “Deep learning and pre- training technology for encrypted traffic classification: A comprehensive review,” Neurocomputing, p. 128444, 2024

work page 2024
[5]

Advancing network security in industrial iot: A deep dive into ai-enabled intrusion detection systems,

M. Shahin, M. Maghanaki, A. Hosseinzadeh, and F. F. Chen, “Advancing network security in industrial iot: A deep dive into ai-enabled intrusion detection systems,” Advanced Engineering Informatics, vol. 62, p. 102685, 2024

work page 2024
[6]

Machine learning for encrypted malicious traffic detection: Approaches, datasets and comparative study,

Z. Wang, K. W. Fok, and V . L. Thing, “Machine learning for encrypted malicious traffic detection: Approaches, datasets and comparative study,” Computers & Security, vol. 113, p. 102542, 2022

work page 2022
[7]

A novel and effective encrypted traffic classification method based on channel attention and deformable convolution,

A. Zou, W. Yang, C. Tang, J. Lu, and J. Guo, “A novel and effective encrypted traffic classification method based on channel attention and deformable convolution,” Computers and Electrical Engineering, vol. 118, p. 109406, 2024

work page 2024

Show all 31 references

[8]

End-to-end encrypted traffic classification with one-dimensional convolution neural networks,

W. Wang, M. Zhu, J. Wang, X. Zeng, and Z. Yang, “End-to-end encrypted traffic classification with one-dimensional convolution neural networks,” in 2017 IEEE international conference on intelligence and security informatics (ISI). IEEE, 2017, pp. 43–48

2017
[9]

Deep packet: A novel approach for encrypted traffic classification using deep learning,

M. Lotfollahi, M. Jafari Siavoshani, R. Shirali Hossein Zade, and M. Saberian, “Deep packet: A novel approach for encrypted traffic classification using deep learning,” Soft Computing, vol. 24, no. 3, pp. 1999–2012, 2020

1999
[10]

Neural architecture search: Insights from 1000 papers,

C. White, M. Safari, R. Sukthanker, B. Ru, T. Elsken, A. Zela, D. Dey, and F. Hutter, “Neural architecture search: Insights from 1000 papers,” arXiv preprint arXiv:2301.08727, 2023

2023 arXiv
[11]

Neural architecture search benchmarks: Insights and survey,

K. T. Chitty-Venkata, M. Emani, V . Vishwanath, and A. K. Somani, “Neural architecture search benchmarks: Insights and survey,” IEEE Access, vol. 11, pp. 25 217–25 236, 2023

2023
[12]

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
[13]

(2022) Hongke sharing — what is deep packet inspection (dpi)? (chinese)

HongKe. (2022) Hongke sharing — what is deep packet inspection (dpi)? (chinese). (accessed 09 October 2024). [Online]. Available: https://zhuanlan.zhihu.com/p/572823255

2022
[14]

Retracted: Flow online identification method for the encrypted skype,

S. Dong and R. Jain, “Retracted: Flow online identification method for the encrypted skype,” 2019

2019
[15]

Random forest based traffic classification method in sdn,

Y . Zhai and X. Zheng, “Random forest based traffic classification method in sdn,” in 2018 international conference on cloud computing, big data and blockchain (ICCBB). IEEE, 2018, pp. 1–5

2018
[16]

Iclstm: encrypted traffic service identification based on inception-lstm neural network,

B. Lu, N. Luktarhan, C. Ding, and W. Zhang, “Iclstm: encrypted traffic service identification based on inception-lstm neural network,” Symmetry, vol. 13, no. 6, p. 1080, 2021

2021
[17]

A session- packets-based encrypted traffic classification using capsule neural net- works,

S. Cui, B. Jiang, Z. Cai, Z. Lu, S. Liu, and J. Liu, “A session- packets-based encrypted traffic classification using capsule neural net- works,” in Proc. 2019 IEEE 21st Int. Conf. High Perform. Comput. Commun.; 17th Int. Conf. Smart City; 5th Int. Conf. Data Sci. Syst. (HPCC/...

2019
[18]

Network traffic classification model based on attention mechanism and spatiotemporal features,

F. Hu, S. Zhang, X. Lin, L. Wu, N. Liao, and Y . Song, “Network traffic classification model based on attention mechanism and spatiotemporal features,” EURASIP J. Inf. Secur., vol. 2023, no. 1, p. 6, 2023

2023
[19]

Neural architecture search with reinforcement learning,

B. Zoph, “Neural architecture search with reinforcement learning,” arXiv preprint arXiv:1611.01578, 2016

2016 arXiv
[20]

Darts: Differentiable architecture search,

H. Liu, K. Simonyan, and Y . Yang, “Darts: Differentiable architecture search,” arXiv preprint arXiv:1806.09055, 2018

2018 arXiv
[21]

A comprehensive survey on hardware-aware neural architecture search,

H. Benmeziane, K. E. Maghraoui, H. Ouarnoughi, S. Niar, M. Wistuba, and N. Wang, “A comprehensive survey on hardware-aware neural architecture search,” arXiv preprint arXiv:2101.09336, 2021

2021 arXiv
[22]

An afford- able hardware-aware neural architecture search for deploying convolu- tional neural networks on ultra-low-power computing platforms,

A. M. Garavagno, E. Ragusa, A. Frisoli, and P. Gastaldo, “An afford- able hardware-aware neural architecture search for deploying convolu- tional neural networks on ultra-low-power computing platforms,” IEEE Sensors Letters, 2024

2024
[23]

Multi-objective hardware-aware neural architecture search using hardware cost diversity,

N. Sinha, P. Rostami, A. El Rahman Shabayek, A. Kacem, and D. Aouada, “Multi-objective hardware-aware neural architecture search using hardware cost diversity,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024, pp. 8032–8039

2024
[24]

Combining com- pressed sensing and neural architecture search for sensor-near vibration diagnostics,

E. Ragusa, F. Zonzini, P. Gastaldo, and L. De Marchi, “Combining com- pressed sensing and neural architecture search for sensor-near vibration diagnostics,” IEEE Transactions on Industrial Informatics, 2024

2024
[25]

Compression- accuracy co-optimization through hardware-aware neural architecture search for vibration damage detection,

E. Ragusa, F. Zonzini, L. De Marchi, and R. Zunino, “Compression- accuracy co-optimization through hardware-aware neural architecture search for vibration damage detection,” IEEE Internet of Things J., 2024

2024
[26]

Tiny neural net- works for session-level traffic classification,

A. Chehade, E. Ragusa, P. Gastaldo, and R. Zunino, “Tiny neural net- works for session-level traffic classification,” in International Conference on Applications in Electronics Pervading Industry, Environment and Society. Springer, 2024, pp. 347–354

2024
[27]

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,” in 2017 International conference on information networking (ICOIN). IEEE, 2017, pp. 712–717

2017
[29]

Identification of encrypted traffic through attention mechanism based long short term memory,

H. Yao, C. Liu, P. Zhang, S. Wu, C. Jiang, and S. Yu, “Identification of encrypted traffic through attention mechanism based long short term memory,” IEEE trans. big data, vol. 8, no. 1, pp. 241–252, 2019

2019
[30]

Encrypted traffic classification based on text convolution neural networks,

M. Song, J. Ran, and S. Li, “Encrypted traffic classification based on text convolution neural networks,” in 2019 IEEE 7th International Conference on Computer Science and Network Technology (ICCSNT). IEEE, 2019, pp. 432–436

2019
[31]

An encrypted traffic classification framework based on convolutional neural networks and stacked autoencoders,

M. Wang, K. Zheng, D. Luo, Y . Yang, and X. Wang, “An encrypted traffic classification framework based on convolutional neural networks and stacked autoencoders,” in 2020 IEEE 6th International Conference on Computer and Communications (ICCC). IEEE, 2020, pp. 634–641

2020

Pith tools

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

[1] [28]

Centime: A direct comprehensive traffic features extraction for en- crypted traffic classification,

W. Maonan, Z. Kangfeng, X. Ning, Y . Yanqing, and W. Xiujuan, “Centime: A direct comprehensive traffic features extraction for en- crypted traffic classification,” in Proc. 2021 IEEE 6th Int. Conf. Comput. Commun. Syst. (ICCCS). IEEE, 2021, pp. 490–498

work page 2021

[2] [1]

A novel traffic classification ap- proach by employing deep learning on software-defined networking,

D. Nu ˜nez-Agurto, W. Fuertes, L. Marrone, E. Benavides-Astudillo, C. Coronel-Guerrero, and F. Perez, “A novel traffic classification ap- proach by employing deep learning on software-defined networking,” Future Internet, vol. 16, no. 5, p. 153, 2024

work page 2024

[3] [2]

Glads: A global-local attention data selection model for multimodal multitask encrypted traffic classification of iot,

J. Dai, X. Xu, and F. Xiao, “Glads: A global-local attention data selection model for multimodal multitask encrypted traffic classification of iot,” Computer Networks, vol. 225, p. 109652, 2023

work page 2023

[4] [3]

Cbs: A deep learning approach for encrypted traffic classification with mixed spatio- temporal and statistical features,

M. Seydali, F. Khunjush, B. Akbari, and J. Dogani, “Cbs: A deep learning approach for encrypted traffic classification with mixed spatio- temporal and statistical features,” IEEE Access, 2023

work page 2023

[5] [4]

Deep learning and pre- training technology for encrypted traffic classification: A comprehensive review,

W. Dong, J. Yu, X. Lin, G. Gou, and G. Xiong, “Deep learning and pre- training technology for encrypted traffic classification: A comprehensive review,” Neurocomputing, p. 128444, 2024

work page 2024

[6] [5]

Advancing network security in industrial iot: A deep dive into ai-enabled intrusion detection systems,

M. Shahin, M. Maghanaki, A. Hosseinzadeh, and F. F. Chen, “Advancing network security in industrial iot: A deep dive into ai-enabled intrusion detection systems,” Advanced Engineering Informatics, vol. 62, p. 102685, 2024

work page 2024

[7] [6]

Machine learning for encrypted malicious traffic detection: Approaches, datasets and comparative study,

Z. Wang, K. W. Fok, and V . L. Thing, “Machine learning for encrypted malicious traffic detection: Approaches, datasets and comparative study,” Computers & Security, vol. 113, p. 102542, 2022

work page 2022

[8] [7]

A novel and effective encrypted traffic classification method based on channel attention and deformable convolution,

A. Zou, W. Yang, C. Tang, J. Lu, and J. Guo, “A novel and effective encrypted traffic classification method based on channel attention and deformable convolution,” Computers and Electrical Engineering, vol. 118, p. 109406, 2024

work page 2024