Pith. sign in

REVIEW 4 major objections 5 minor 40 references

Learning in Multiple Spaces: Few-Shot Network Attack Detection with Metric-Fused Prototypical Networks

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Pith's one-line read The paper claims that a prototypical few-shot network fusing Euclidean, cosine, Chebyshev, and Wasserstein distances outperforms single-metric baselines for detecting rare and novel network attacks on three intrusion datasets.

desk verdict A plausible multi-metric prototypical-network extension whose zero-day claim is unsupported because the evaluation never holds out attack classes. read the letter →

arxiv 2501.00050 v1 pith:D6FAPESE submitted 2024-12-28 cs.CR cs.LG

classification cs.CRcs.LG
keywords few-shotlearningnetworkintrusiondetectionprototypicalnetworksmetriczero-dayattackPolyakaveragingepisodictrainingdatasets
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that a few-shot network intrusion detector performs better when it compares each sample to class prototypes in several metric spaces at once—Euclidean, cosine, Chebyshev, and Wasserstein—rather than in a single space. Using only 200 training samples per dataset and 40 random seeds, the reported experiments show the multi-space approach improving balanced accuracy, validation F1-score, and area under the precision-recall curve (AUPRC) on CICEVSE2024, CICIDS2017, and CICIoV2024 relative to a single-metric prototypical baseline. The practical point is that an intrusion detection system could begin catching a rare or newly appearing attack type with very few labeled flows, which is exactly the setting where conventional detectors fail.

What carries the argument

The load-bearing object is the fused distance $D(x,c_k)=\sum_{m\in M} w_m \hat{d}_m(x,c_k)$, where $M$ contains Euclidean, cosine, Chebyshev, and Wasserstein distances, each z-score normalized and clipped so no single metric dominates, with weights constrained by $\sum_{m\in M}w_m=1$, $w_m\ge 0$. The class prototype $c_k$ is the mean embedding of the support set computed with an exponential moving average of the model parameters (Polyak averaging), which stabilizes prototype estimates across episodes. Balanced episodic sampling then builds support and query sets with controlled repetition for rare classes, and the network is trained with a sigmoid cross-entropy loss on the fused distances.

What would settle it

Hold out entire attack classes in CICIDS2017, train MSPL and the single-metric baseline on the remaining classes with 200 samples, and compare AUPRC on the held-out classes only; if the gains vanish, the zero-day claim is unsupported. As a second check, rerun all three datasets with one fixed set of metric weights across datasets to see whether the reported improvements survive without per-dataset tuning.

Watch

Extended reading notes

Core claim

On the paper's own terms, the core discovery is that fusing four normalized distance metrics in a prototypical network consistently beats the Euclidean-only version in a 200-sample few-shot regime. The reported gains are a balanced accuracy increase from roughly 0.79 to 0.81–0.82 on CICEVSE2024, from 0.85–0.86 to 0.87–0.88 on CICIDS2017, and from 0.96–0.97 to 0.97–0.98 on CICIoV2024, with AUPRC improving most on the datasets where the baseline scored lowest. The authors read these results as evidence that metric complementarity, Polyak-averaged prototypes, and balanced episodic training let the framework adapt to low-profile and zero-day attacks.

Load-bearing premise

The zero-day claim rests on validation episodes actually containing attack classes never seen during training, and on the per-dataset metric weights being fixed before inspecting validation performance; the paper does not explicitly document either condition.

Editorial extensions

If this is right

  • A deployed detector could be updated to recognize a new attack class from a few hundred labeled flows by adding the multi-metric scoring layer to an existing prototypical network.
  • Datasets with strong class imbalance should show the largest AUPRC gains, since the method concentrates its improvement on low-profile classes.
  • The same metric-fusion recipe applies to other few-shot metric-based classifiers, not only intrusion detection.
  • The per-dataset choice among bi-, tri-, and quad-metric configurations becomes a practical tuning decision for operators.
  • Polyak-averaged prototypes can be used as a general stabilizer for episodic few-shot training, independent of the metric blend.

Reading between the lines

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

  • The paper does not explicitly state whether validation episodes draw from attack classes held out of training, so the reported numbers should be read as evidence about scarce-data classification of seen classes until a class-split experiment is reported.
  • The per-dataset metric weights in Table 1 vary without a documented selection procedure; a fair comparison would fix the weights before seeing validation results or use nested cross-validation.
  • A direct test of the zero-day claim would train on some attack families and evaluate only on families never seen in training, comparing MSPL's AUPRC with the single-metric baseline on those families.
  • If equal fixed weights across all datasets still produce gains, the benefit is likely from metric diversity itself rather than from dataset-specific tuning.
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 / 5 minor

Summary. The manuscript proposes Multi-Space Prototypical Learning (MSPL), a few-shot network intrusion detection method that combines Euclidean, cosine, Chebyshev, and Wasserstein distances in a prototypical network, with constrained metric weights, Polyak averaging of model parameters, and balanced episodic training. The authors evaluate on three CIC datasets (CICEVSE2024, CICIDS2017, CICIoV2024) with 200 training samples and 40 seeds, reporting balanced accuracy, validation F1, and AUPRC. They report gains over a single-metric Euclidean prototypical baseline and over simple non-deep baselines, and they interpret these gains as evidence for effective zero-day attack detection.

Significance. The underlying idea of fusing several distance metrics in a prototypical network is plausible, and the paper has some strengths: it uses three public benchmark datasets, reports means and standard deviations over 40 seeds, and makes the algorithmic procedure explicit in Algorithm 1. However, the claimed zero-day generalization is not established by the experiments because no attack class is held out, and the per-dataset choice of metric weights creates a post-hoc selection risk. If the authors add a class-disjoint evaluation protocol, fix the weighting rule, and compare with existing few-shot IDS methods, the empirical claims could be meaningfully assessed. As it stands, the evidence supports only a modest claim about few-shot classification of known attack classes with a tuned multi-metric prototype model.

major comments (4)
  1. [Evaluation, Algorithm 1] The central claim that MSPL 'effectively adapts to rare and zero-day attacks' is not supported by the evaluation, because Algorithm 1 draws all training and validation episodes from the same stratified sample (Xtrain, Ytrain) of D and never holds out an entire attack class. In standard prototypical-network few-shot evaluation, zero-day generalization requires training on base classes and testing on held-out novel classes; without such a class-disjoint split, the reported balanced accuracy, F1, and AUPRC measure few-shot classification of classes already seen in training. Please either add a class-disjoint split (e.g., leave one or more attack classes out of all training episodes and test on them) or revise the abstract and conclusion to claim only adaptation to rare classes within the known label set.
  2. [Table 1, Polyak Averaging Effectiveness] The results in Table 1 contradict the claim that Polyak averaging stabilizes or improves performance. On CICIoV2024, the Polyak-off row has higher balanced accuracy (0.9813 vs. 0.9718), higher F1 (0.9718 vs. 0.9404), and higher AUPRC (0.9506 vs. 0.8992) than the Polyak-on row; on CICEVSE2024 the Polyak-off row is also uniformly better; on CICIDS2017 the comparison is mixed. The paper does not state which configuration defines 'MSPL' or how the Polyak setting was selected. This must be clarified, and the claim about Polyak averaging should be limited to the configurations where it actually helps, or supported by a principled selection rule.
  3. [Table 1, Metric Weights] The metric weights w_m are chosen separately for each dataset and even separately for the Polyak-on and Polyak-off variants, with no described selection procedure. Because Table 1 reports the best configuration per dataset, the reported MSPL gains may reflect post-hoc configuration choice rather than the method's intrinsic advantage. To make the comparison fair, the authors should specify a single weighting rule, or a validation-based selection procedure applied identically to all variants, and report results under that rule.
  4. [Evaluation, Comparison] The evaluation uses only a Euclidean-only baseline and unspecified non-deep models as comparators, and all numbers are validation metrics (Algorithm 1, line 23); there is no separate test set and no comparison with existing few-shot intrusion detection methods such as FS-IDS, the supervised-autoencoder approach, or the dual-space prototypical network this work builds on. Without a held-out test set and published baselines, the statement that MSPL 'outperforms traditional approaches' is too strong. Please add a test set disjoint from validation and compare with at least one or two established few-shot IDS methods under the same protocol.
minor comments (5)
  1. [Throughout] The dataset is referred to as both CICEVSE2024 and CICEVSE; use a single consistent name throughout the paper.
  2. [Methodology, Wasserstein distance] The Wasserstein distance is under-specified: it is unclear whether F_x and F_ck denote empirical CDFs over the dimensions of a single embedding or over a batch of samples, and how the expectation is computed in practice.
  3. [Problem Formulation] The task is called multi-label, but the episodic C-way K-shot setup and the prototype/query construction describe single-label classification; please clarify the relationship between the two formulations.
  4. [Algorithm 1] Algorithm 1 says 'Validate model and save if improved' without specifying the validation metric, the validation set construction, or the early-stopping criterion; this information is needed for reproducibility.
  5. [Evaluation, Non-DL baselines] The non-DL baselines (Logistic Regression, Random Forest, Gradient Boosting) are not described, nor are their hyperparameters or feature inputs; please provide experimental details.

Circularity Check

2 steps flagged · score 6.0 of 10

The zero-day claim renames in-distribution validation scores as novel-class generalization, and the per-dataset metric weights are post-hoc selections reported as a fixed MSPL prediction.

  1. renaming known result [Abstract; Algorithm 1 lines 2-6 and 23; Evaluation (FSL adaptation)]
    "By leveraging Polyak-averaged prototype generation, the framework stabilizes the learning process and effectively adapts to rare and zero-day attacks. ... (Xtrain, Ytrain) ← sample(D); E ←CreateEpisodes(Xtrain, Ytrain, Ne, Ns, Nq); ... Validate model and save if improved."

    Algorithm 1 draws a single stratified sample from D, builds every training and validation episode from that same pool, and saves the model when validation improves on that same pool. No attack class is ever held out of training. Consequently, the reported balanced accuracy, validation F1, and AUPRC are, by construction, few-shot classification scores on classes already seen during training. Calling these results 'zero-day attack detection' renames an in-distribution validation score as a novel-class generalization result; the central empirical claim reduces to the same known-class validation numbers used for early stopping and model selection.

  2. fitted input called prediction [Table 1; Metric Space Contribution; Key Observations]
    "On the CICEVSE2024 dataset, the tri-metric approach enhanced pattern recognition by leveraging distinct geometric and directional properties (Table 1, Metric Weights column). For CICIDS2017, the bi-metric design effectively reduced overfitting to specific attack types. ... On CICIoV2024, a quad-metric configuration (Euclidean, Chebyshev, Wasserstein, Cosine) provided balanced contributions from all metrics."

    The metric weights in Table 1 are not learned or derived by a stated procedure; they are per-dataset and per-variant hand choices, and the bolded 'MSPL' performance is the best of these configurations. For example, CICIDS2017 uses (1/3,1/3,0,1/3) without Polyak but (1/2,0,0,1/2) with Polyak, while CICEVSE and CICIoV use different tri- and quad-metric splits. The paper does not describe a validation-based selection protocol, so the reported MSPL 'prediction' of superiority is partly constructed by choosing the input weights after inspecting validation performance, then presenting the best validation numbers as the framework's forecast.

full rationale

The two flagged issues are not equation-level derivation loops, but they do make the headline 'zero-day' claim reduce to in-distribution validation. Algorithm 1 never defines a class-disjoint split: (Xtrain,Ytrain) is a single stratified sample of D, all episodes are built from it, and the model is saved on validation from the same pool. The reported metrics are therefore known-class few-shot scores, and the abstract's 'zero-day' statement is a renaming of those scores rather than a measured novel-class result. Separately, Table 1's metric weights vary per dataset and per Polyak setting with no stated selection rule, and the best variant is bolded as MSPL; this is post-hoc configuration choice on validation. These are partial circularities because the claimed predictions are built from the same validation numbers used to choose the model, but the Euclidean baseline and three public benchmark datasets provide independent empirical content, and the self-citation to Martinez et al. (2024) is acknowledged prior work rather than a load-bearing uniqueness theorem. I therefore score 6 rather than 8-10.

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

The framework adds no new physical entities. The main unpaid assumptions are evaluation-based: that validation episodes represent zero-day attacks, and that the hand-picked metric weights are a general property of the method. All key hyperparameters are undisclosed.

free parameters (5)
  • Metric weights per dataset = CICEVSE: (1/3,1/3,0,1/3); CICIDS no-Polyak: (1/3,1/3,0,1/3); CICIDS Polyak: (1/2,0,0,1/2); CICIoV: (1/4,1/4,1/4,1/4)
    Weights are chosen by hand per dataset and per variant, not learned or fixed, and appear selected after observing performance.
  • Z-score clipping threshold gamma = not reported
    Used in normalized distance computation; the value is not given.
  • Polyak averaging decay beta = not reported
    EMA decay rate for model parameters is never specified.
  • Support and query sizes and episode count = not reported
    The C-way K-shot configuration is never given, yet the table results depend on it.
  • Embedding network architecture and optimizer settings = not reported
    No details of f_theta, learning rate, optimizer, or gradient clipping are provided.
assumptions (5)
  • domain assumption Each embedding vector and prototype vector can be treated as a probability distribution for Wasserstein distance.
    The formula for dW uses inverse CDFs of x and c_k without describing how a single d-dimensional vector defines a distribution.
  • domain assumption Distance normalization via z-score with clipping preserves the information needed for weighted classification.
    The metric integration section introduces mu_m and sigma_m but does not specify the scope of their computation or justify the clipping.
  • domain assumption Episodic training with stratified sampling and repetition sampling generalizes to unseen attack classes.
    The Episodic Learning Framework only ensures balanced classes within training episodes; no held-out class experiment demonstrates zero-day generalization.
  • domain assumption Sigmoid over negative combined distance is a valid multi-label classification objective.
    Equation (1) defines the loss without comparing to softmax over prototypes or other calibrated objectives.
  • domain assumption The benchmark datasets are correctly labeled and the selected features are suitable for the few-shot setup.
    No preprocessing or feature selection details are given; dataset ground truth is taken from prior work.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning in Multiple Spaces: Few-Shot Network Attack Detection with Metric-Fused Prototypical Networks." pith.science (2026). https://pith.science/paper/D6FAPESE

@misc{pith2026250100050,
  author       = {Pith},
  title        = {Pith review of: Learning in Multiple Spaces: Few-Shot Network Attack Detection with Metric-Fused Prototypical Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D6FAPESE}},
  note         = {Machine review of arXiv:2501.00050}
}
read the original abstract

Network intrusion detection systems face significant challenges in identifying emerging attack patterns, especially when limited data samples are available. To address this, we propose a novel Multi-Space Prototypical Learning (MSPL) framework tailored for few-shot attack detection. The framework operates across multiple metric spaces-Euclidean, Cosine, Chebyshev, and Wasserstein distances-integrated through a constrained weighting scheme to enhance embedding robustness and improve pattern recognition. By leveraging Polyak-averaged prototype generation, the framework stabilizes the learning process and effectively adapts to rare and zero-day attacks. Additionally, an episodic training paradigm ensures balanced representation across diverse attack classes, enabling robust generalization. Experimental results on benchmark datasets demonstrate that MSPL outperforms traditional approaches in detecting low-profile and novel attack types, establishing it as a robust solution for zero-day attack detection.

Figures

Figures reproduced from arXiv: 2501.00050 by the authors.

Figure 1
Figure 1. Balanced accuracy comparison for baselines and MSPL approaches across CICEVSE, CICIDS2017, and CICIoV2024 [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. F1-score comparison for baselines and MSPL approaches across CICEVSE, CICIDS2017, and CICIoV2024 datasets [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. AUPRC comparison for baselines and MSPL experiments across CICEVSE, CICIDS2017, and CICIoV2024 datasets. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 37 canonical work pages

  1. [1]

    Aharon, U.; Dubin, R.; Dvir, A.; and Hajaj, C. 2024 a . Few-Shot API Attack Anomaly Detection in a Classification-by-Retrieval Framework. arXiv preprint arXiv:2405.11247

  2. [2]

    Aharon, U.; Marbel, R.; Dubin, R.; Dvir, A.; and Hajaj, C. 2024 b . Few-Shot API Attack Detection: Overcoming Data Scarcity with GAN-Inspired Learning. arXiv preprint arXiv:2405.11258

  3. [3]

    Bilot, T.; El Madhoun, N.; Al Agha, K.; and Zouaoui, A. 2024. Few Edges are Enough: Few-Shot Network Attack Detection with Graph Neural Networks. In International Workshop on Security, 257--276. Springer

  4. [4]

    D.; Ghorbani, A

    Buedi, E. D.; Ghorbani, A. A.; Dadkhah, S.; and Ferreira, R. 2024. Enhancing EV Charging Station Security Using A Multi-dimensional Dataset: CICEVSE2024. In Submitted to ESORICS 2024 Conference. Submitted

  5. [5]

    Chowdhury, M. M. U.; Hammond, F.; Konowicz, G.; Xin, C.; Wu, H.; and Li, J. 2017. A few-shot deep learning approach for improved intrusion detection. In 2017 IEEE 8th Annual Ubiquitous Computing, Electronics and Mobile Communication Conference (UEMCON), 456--462. IEEE

  6. [6]

    K.; and Angin, P

    Demirpolat, A.; Sarica, A. K.; and Angin, P. 2021. Prot \'E dge: a few-shot ensemble learning approach to software-defined networking-assisted edge security. Transactions on Emerging Telecommunications Technologies, 32(6): e4138

  7. [7]

    Du, L.; Gu, Z.; Wang, Y.; Wang, L.; and Jia, Y. 2023. A Few-Shot Class-Incremental Learning Method for Network Intrusion Detection. IEEE Transactions on Network and Service Management

  8. [8]

    Duan, R.; Li, D.; Tong, Q.; Yang, T.; Liu, X.; and Liu, X. 2021. A Survey of Few-Shot Learning: An Effective Method for Intrusion Detection. Security and Communication Networks, 2021(1): 4259629

Show all 40 references
  1. [9]

    Gel, B.; Li, B.; Mou, X.; Zhao, J.; and Liu, X. 2024. A Few-Shot Network Flow Attack Classification via Graph Contrastive Learning. In 2024 IEEE 11th International Conference on Cyber Security and Cloud Computing (CSCloud), 30--35. IEEE

  2. [10]

    Guo, Y. 2023. A review of Machine Learning-based zero-day attack detection: Challenges and future directions. Computer Communications, 198: 175--185

  3. [11]

    K.; Niu, W.; Zhang, X.; and Li, F

    He, J.; Yao, L.; Li, X.; Khan, M. K.; Niu, W.; Zhang, X.; and Li, F. 2024. Model-agnostic generation-enhanced technology for few-shot intrusion detection. Applied Intelligence, 54(4): 3181--3204

  4. [12]

    He, M.; Wang, X.; Zhou, J.; Xi, Y.; Jin, L.; and Wang, X. 2021. Deep-Feature-Based Autoencoder Network for Few-Shot Malicious Traffic Detection. Security and Communication Networks, 2021(1): 6659022

  5. [13]

    S.; Abdurrahman, U

    Iliyasu, A. S.; Abdurrahman, U. A.; and Zheng, L. 2022. Few-shot network intrusion detection using discriminative representation learning with supervised autoencoder. Applied Sciences, 12(5): 2351

  6. [14]

    Kale, R.; and Thing, V. L. 2023. Few-shot weakly-supervised cybersecurity anomaly detection. Computers & Security, 130: 103194

  7. [15]

    Liang, W.; Hu, Y.; Zhou, X.; Pan, Y.; Kevin, I.; and Wang, K. 2021. Variational few-shot learning for microservice-oriented intrusion detection in distributed industrial IoT. IEEE Transactions on Industrial Informatics, 18(8): 5087--5095

  8. [16]

    Liu, C.; Li, B.; Zhao, J.; Zhen, Z.; Liu, X.; and Zhang, Q. 2022. FewM-HGCL: Few-shot malware variants detection via heterogeneous graph contrastive learning. IEEE Transactions on Dependable and Secure Computing

  9. [17]

    Lu, C.; Wang, X.; Yang, A.; Liu, Y.; and Dong, Z. 2023. A Few-Shot-Based Model-Agnostic Meta-Learning for Intrusion Detection in Security of Internet of Things. IEEE Internet of Things Journal, 10(24): 21309--21321

  10. [18]

    Ma, W.; Ma, L.; Li, K.; and Guo, J. 2023. Few-shot IoT attack detection based on SSDSAE and adaptive loss weighted meta residual network. Information Fusion, 98: 101853

  11. [19]

    Martinez, F.; Mapkar, M.; Alfatemi, A.; Rahouti, M.; Xin, Y.; Xiong, K.; and Ghani, N. 2024. Redefining DDoS Attack Detection Using A Dual-Space Prototypical Network-Based Approach. In 2024 33rd International Conference on Computer Communications and Networks (ICCCN), 1--9

  12. [20]

    Miao, G.; Wu, G.; Zhang, Z.; Tong, Y.; and Lu, B. 2023. Spn: A method of few-shot traffic classification with out-of-distribution detection based on siamese prototypical network. IEEE Access

  13. [21]

    Neto, E. C. P.; Taslimasa, H.; Dadkhah, S.; Iqbal, S.; Xiong, P.; Rahmanb, T.; and Ghorbani, A. A. 2024. CICIoV2024: Advancing Realistic IDS Approaches against DoS and Spoofing Attack in IoV CAN bus. Internet of Things, 101209

  14. [22]

    Ouyang, Y.; Li, B.; Kong, Q.; Song, H.; and Li, T. 2021. FS-IDS: a novel few-shot learning based intrusion detection system for scada networks. In ICC 2021-IEEE International Conference on Communications, 1--6. IEEE

  15. [23]

    Pan, H.; Fang, Y.; Guo, W.; Xu, Y.; and Wang, C. 2024. Few-shot graph classification on cross-site scripting attacks detection. Computers & Security, 140: 103749

  16. [24]

    Parnami, A.; and Lee, M. 2022. Learning from few examples: A summary of approaches to few-shot learning. arXiv preprint arXiv:2203.04291

  17. [25]

    Qin, Y.; Zhao, C.; Zhu, X.; Wang, Z.; Yu, Z.; Fu, T.; Zhou, F.; Shi, J.; and Lei, Z. 2020. Learning meta model for zero-and few-shot face anti-spoofing. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, 11916--11923

  18. [26]

    Rong, C.; Gou, G.; Hou, C.; Li, Z.; Xiong, G.; and Guo, L. 2021. Umvd-fsl: Unseen malware variants detection using few-shot learning. In 2021 international joint conference on neural networks (IJCNN), 1--8. IEEE

  19. [27]

    Sharafaldin, I.; Habibi Lashkari, A.; and Ghorbani, A. A. 2018. Toward Generating a New Intrusion Detection Dataset and Intrusion Traffic Characterization. In 4th International Conference on Information Systems Security and Privacy (ICISSP). Portugal

  20. [28]

    Sun, H.; Huang, Y.; Zhou, C.; Han, L.; Liu, H.; Chen, J.; and Li, X. 2024. Space Decoupled Prototype Learning for Few-Shot Attack Detection in Cyber--Physical Systems. IEEE Transactions on Industrial Informatics

  21. [29]

    T.; Shiraishi, Y.; and Morii, M

    Thein, T. T.; Shiraishi, Y.; and Morii, M. 2023. Few-Shot Learning-Based Malicious IoT Traffic Detection with Prototypical Graph Neural Networks. IEICE TRANSACTIONS on Information and Systems, 106(9): 1480--1489

  22. [30]

    Tian, Y.; Zheng, N.; Chen, X.; and Gao, L. 2021. Wasserstein Metric-Based Location Spoofing Attack Detection in WiFi Positioning Systems. Security and Communication Networks, 2021(1): 8817569

  23. [31]

    P.; Swain, A

    Vijayakanthi, G.; Mohanty, J. P.; Swain, A. K.; and Mahapatra, K. 2021. Differential metric based deep learning methodology for non-profiled Side Channel Analysis. In 2021 IEEE International Symposium on Smart Electronic Systems (iSES), 200--203. IEEE

  24. [32]

    T.; and Ni, L

    Wang, Y.; Yao, Q.; Kwok, J. T.; and Ni, L. M. 2020. Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys (csur), 53(3): 1--34

  25. [33]

    Wang, Z.-M.; Tian, J.-Y.; Qin, J.; Fang, H.; and Chen, L.-M. 2021. A Few-Shot Learning-Based Siamese Capsule Network for Intrusion Detection with Imbalanced Training Data. Computational intelligence and neuroscience, 2021(1): 7126913

  26. [34]

    Wei, W.; Gu, H.; Deng, W.; Xiao, Z.; and Ren, X. 2022. ABL-TC: A lightweight design for network traffic classification empowered by deep learning. Neurocomputing, 489: 333--344

  27. [35]

    Xu, C.; Shen, J.; and Du, X. 2020. A method of few-shot network intrusion detection based on meta-learning framework. IEEE Transactions on Information Forensics and Security, 15: 3540--3552

  28. [36]

    Yang, J.; Li, H.; Shao, S.; Zou, F.; and Wu, Y. 2022. FS-IDS: A framework for intrusion detection based on few-shot learning. Computers & Security, 122: 102899

  29. [37]

    Yu, Y.; and Bian, N. 2020. An intrusion detection method using few-shot learning. IEEE Access, 8: 49730--49740

  30. [38]

    Zhou, X.; Liang, W.; Shimizu, S.; Ma, J.; and Jin, Q. 2020. Siamese neural network based few-shot learning for anomaly detection in industrial cyber-physical systems. IEEE Transactions on Industrial Informatics, 17(8): 5790--5798

  31. [39]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  32. [40]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.