Pith. sign in

REVIEW 4 major objections 4 minor 40 references

This paper claims that treating augmented benign samples as negatives, instead of as positives, makes a self-supervised model learn a single benign-traffic cluster and detect intrusions by distance from one centroid.

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

CLAN clusters genuine benign network flows while repelling augmented copies, then classifies new flows by distance to the cluster centroid; on Lycos2017 it reports the highest mean AUROC among compared SSL and anomaly detectors.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection CLAN's flipped contrastive pairing is a genuine novelty and the Lycos2017 numbers are strong, but the theoretical derivation is shaky and the load-bearing surrogate assumption is untested. the 4 major comments →

arxiv 2509.06550 v1 pith:YB6D74IT submitted 2025-09-08 cs.LG cs.AIcs.CRcs.NI

Contrastive Self-Supervised Network Intrusion Detection using Augmented Negative Pairs

classification cs.LG cs.AIcs.CRcs.NI
keywords self-supervised learningcontrastive learningnetwork intrusion detectionanomaly detectionaugmented negative pairsbenign-only trainingLycos2017
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

CLAN proposes flipping the usual contrastive-learning sign convention for network intrusion detection. Instead of pulling each sample and its augmented views together and pushing other samples apart, it treats augmented samples as stand-ins for malicious traffic and uses other benign samples as the positives. The result, the paper argues, is a latent space where genuine benign traffic collapses into one Gaussian-like cluster while attacks fall outside it, so a single cached centroid can serve as the detector. On Lycos2017 the authors report a mean class-wise AUROC of 0.958591 without any fine-tuning, above the self-supervised and anomaly-detection baselines they compare against, and better few-shot multi-class F1 after fine-tuning in most settings. If correct, the practical draw is that a network can be protected using only benign traffic, at constant inference cost per flow, without labelled attack data.

Core claim

The central claim is that the usual contrastive sign convention is backwards for intrusion detection. Conventionally, an augmented view of a benign sample is pulled toward the sample and other samples are pushed away; CLAN instead pulls different benign samples toward one another and pushes augmented versions away, treating them as surrogates for malicious traffic. Under a Gaussian-latent assumption, the paper derives a batch loss whose optimum is a single compact benign cluster with a well-defined centroid. At inference, the fixed network projects a test flow and compares its distance to the cached centroid, producing a probability of being benign. The paper reports that on Lycos2017 this p

What carries the argument

The load-bearing object is the CLAN loss: a batched contrastive objective with hinge regularisation. For each benign flow, it contains a positive term pulling every other benign flow in the batch closer, and a negative term pushing an augmented version at least a margin distance away. The augmentation is per-feature uniform resampling within a bounded range with a fixed resampling probability, which the paper treats as a surrogate for a potentially malicious class distribution. The loss is derived as maximum-likelihood estimation of a homoskedastic Gaussian latent model, with centroid distances replaced by Monte Carlo estimates. The geometric mean of training latent vectors gives the cached

Load-bearing premise

The entire detector hinges on the surrogate augmentation, uniformly resampling benign features, standing in for real malicious traffic; if that surrogate does not resemble attacks in latent space, the hinge term just pushes away noise rather than intrusions.

What would settle it

A direct test is to train CLAN on benign Lycos2017 flows and measure its AUROC separately on held-out real attack samples and on held-out pure augmentation noise. If the model separates noise better than real attacks, or if its mean class-wise AUROC drops sharply on a second intrusion dataset with different attack statistics, the surrogate assumption is falsified.

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

If this is right

  • Deployed as an anomaly detector, CLAN can label a flow by one distance to a precomputed centroid, making per-flow inference cost constant rather than requiring a nearest-neighbour search over training data.
  • The same pretrained representation transfers to multi-class classification when fine-tuned with few labels per class, outperforming the compared self-supervised baselines in macro-F1 across most reported dataset sizes.
  • Because training uses only benign traffic, a newly deployed network that has not yet observed any attack can still obtain a detector, provided its unlabelled capture is clean.
  • If augmented views are used as negatives, the learned latent geometry is a class-level Gaussian rather than a per-sample mixture, which is exactly the structure a distance-to-centroid classifier needs.

Where Pith is reading between the lines

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

  • The paper never directly measures whether the uniform-resampling surrogate occupies the same latent region as real attacks. That link is the part most likely to be dataset-specific; a straightforward diagnostic would be to compute latent-space distances among the benign centroid, real malicious samples, and pure surrogate samples on a second intrusion dataset.
  • The same flip-of-sign idea should transfer to other single-class or majority-class representation-learning problems: whenever one class is abundant and unlabelled and the rest are rare, augmented versions of the abundant class can serve as synthetic negatives.
  • If the benign-only assumption is violated by polluted training data, the cached centroid shifts and the O(1) inference scheme has no built-in correction. The paper notes this limitation; robustness testing under contamination would determine how much pollution the boundary tolerates.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes CLAN, a self-supervised contrastive learning objective for network intrusion detection trained exclusively on benign traffic. Unlike standard contrastive methods that treat augmented views as positive pairs, CLAN treats augmented samples as negative pairs, supposedly representing a surrogate malicious distribution, while other benign samples serve as positives. The resulting latent space is meant to place benign traffic in a single cluster around a centroid; inference labels a test point by the distance to that cached centroid. Experiments on Lycos2017 compare CLAN to several SSL baselines and anomaly detectors for binary AUROC, and to SSL baselines for fine-tuned multiclass F1 with limited labels. The paper claims improved accuracy and O(1) inference.

Significance. If the central mechanism works, the idea is attractive and practically relevant: it flips the usual role of augmentation, learns a holistic benign distribution rather than per-sample distributions, and permits cheap centroid-based inference. The empirical protocol is a strength: random search with 5-fold cross-validation, retraining on the full benign training split, and a held-out test set, with comparisons against a broad set of SSL and anomaly-detection baselines. However, the theoretical derivation has serious gaps and the key assumption that uniform feature resampling represents real malicious traffic is unvalidated. The significance is therefore conditional: the paper is a promising empirical proposal, but the probabilistic framing and the surrogate assumption need substantial revision before the claims are trustworthy.

major comments (4)
  1. [Section III-A, Eqs. (1)-(3)] The step from Eq. (1) to Eq. (2) is not justified. With Gaussian class-conditional densities and equal priors, P(y=0|x) is N(z;mu0,sigma^2 I) / sum_c N(z;mu_c,sigma^2 I), not the two-term ratio with a single mu_psi. The introduction of a single 'malicious' mean mu_psi is an additional assumption that is not stated or derived. Eq. (3) then sums over Nc-1 negative means, which is inconsistent with Eq. (2), and the negative-log-likelihood simplification drops the variance and prior terms. Consequently, the claim that CLAN is a maximum-likelihood estimator is not established; Eq. (6) is best described as a heuristic contrastive loss.
  2. [Section III-A, surrogate distribution psi(x)] The load-bearing assumption that the surrogate psi(x) (per-feature uniform resampling in [-b,b]) represents the malicious distribution is asserted in one sentence: 'It assumed that an even number of samples of each class, both benign and malicious, are drawn from this distribution.' No evidence is given that uniform feature noise overlaps with real attacks such as portscan, DDoS, or web attacks, either in feature space or after the learned map. The hinge term in Eq. (6) only repels augmented samples; whether real malicious samples are also repelled is untested. A second dataset, an ablation, or a latent-space analysis showing that real attacks map near the augmented negatives is needed to support the binary-detection claim.
  3. [Section III-A, final paragraph; Section V] The experiments use cosine distance, but the derivation in Eqs. (1)-(6) assumes squared Euclidean distance under a homoskedastic Gaussian model. The paper states, 'It can be shown that using the cosine distance metric optimises a similar objective while replacing the Gaussian assumption with a von Mises-Fisher distribution assumption,' but no derivation, citation, or explicit form of the cosine-based loss is provided. Since the implemented objective differs from the derived one, the theoretical motivation does not directly support the reported results. This needs to be made rigorous or the theoretical claims should be scaled back.
  4. [Tables I and II] The main quantitative claims are based on single point estimates with no confidence intervals, standard deviations, or number of seeds. In Table III the difference between CLAN and BYOL at 256 samples/class is 0.002 (0.655416 vs. 0.657744), almost certainly within run-to-run noise, yet the text calls it an 'isolated occurrence.' Without variance estimates or multiple seeds, the 'outperforms' claims across Tables I-III are not statistically supported. Reporting means and standard deviations over at least 5 seeds for the main tables is necessary.
minor comments (4)
  1. [Section III-A, Eq. (4)] The statement that d(z, mu_i) is 'proportional' to E[d(z, z')] is imprecise: for an isotropic Gaussian, E||z-z'||^2 = ||z-mu||^2 + q sigma^2, i.e., an additive constant rather than a purely proportional relationship. Since the constant is independent of z, the optimization use is fine, but the wording should be corrected.
  2. [Section III-B, Eqs. (1) and (9)] In Eq. (1), Z(x) is the x-dependent partition function over classes; in Eq. (9), Z is redefined as a constant hyperparameter. If Z is a tunable scalar, P(y=0|x) is not a calibrated probability and the threshold of 0.5 in Eq. (10) is arbitrary. This should be stated explicitly.
  3. [Section IV] The complexity comparison O(1) vs O(Ntrain) assumes that existing SSL methods require a nearest-neighbour search at inference. This is plausible but should be supported by describing the exact binary decision rule used for each baseline in Tables I-II, or by a citation.
  4. [Throughout] Typos and minor wording: 'wholistic' should be 'holistic'; 'datset' -> 'dataset'; 'Isolation Forrest' -> 'Isolation Forest'; 'Neural Newtwork' -> 'Neural Network'; 'preformed' -> 'performed'; 'It assumed' -> 'It is assumed'; 'will evaluated' -> 'will be evaluated'. Eq. (2) contains a stray 'I' in the denominator. Table III values are missing spaces and should be aligned; standard deviations should be added.

Circularity Check

0 steps flagged

No significant circularity: CLAN's derivation is self-contained and tested on held-out data; self-citations are background only.

full rationale

The paper's central derivation (Eqs. 1-6) is self-contained: it starts from a Gaussian latent-class assumption, uses Monte Carlo estimation (Eq. 4) to replace centroid distances with sample distances, and defines a surrogate ψ(x) for the unavailable malicious class. The final loss (Eq. 6) directly optimizes that objective, and the binary decision rule (Eq. 9) uses a centroid computed from benign training data, independent of test labels. The chief external input is the surrogate augmentation; whether uniform resampling matches real attack traffic is a modeling assumption, not a circularity, and the paper evaluates it on a held-out Lycos2017 test set. The self-citations ([1], [3]) appear only as background references for general ML/SVM context and are not load-bearing for the proposed method or its claims. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from prior authors, and no known result is merely relabeled. Therefore, no step in the derivation reduces by construction to its own inputs.

Axiom & Free-Parameter Ledger

6 free parameters · 7 axioms · 1 invented entities

The CLAN objective (Eq. 6) is assembled from a Gaussian generative story (Eqs. 1-3), a Monte Carlo centroid trick (Eq. 4), an invented surrogate class (ψ), and a hinge margin. None of the constants involved is derived; p_resample, b, m, Z, architecture, and the distance metric are tuned or chosen on validation data and their final values are not reported. The honest contribution is the objective plus its single-dataset evaluation, not a parameter-free derivation.

free parameters (6)
  • p_resample (augmentation probability) = tuned via 200-iteration random search; final value not reported
    Controls how aggressively benign flows are resampled into the surrogate malicious distribution; Section III-A, V-A.
  • b (resampling range) = tuned; value not reported
    Uniform resampling range [-b, b] defining ψ; Section III-A, V-A.
  • m (hinge margin) = tuned; value not reported
    Margin in max(0, m - d(...)) of Eq. 6.
  • Z (inference threshold) = constant hyperparameter; value not reported
    Eq. 9 calls Z a partition function but it is set as a hyperparameter controlling FPR/recall; with threshold 0.5 it defines the distance boundary log(2Z).
  • Architecture and optimization hyperparameters (width, depth, d_model, d_head, learning rate, batch size, weight decay) = selected by random search; values not reported
    Section V-A; the tuned configuration used for the reported tables is not disclosed.
  • Distance metric (cosine vs Euclidean) = cosine chosen after validation
    Section III-A: cosine 'was found to improve performance'.
axioms (7)
  • domain assumption Each class in latent space follows a homoskedastic Gaussian distribution (distinct mean μ_i, shared isotropic covariance σ²I)
    Section III-A, before Eq. 2; the whole NLL derivation (Eqs. 2-3) depends on this generative form.
  • standard math Distance from a sample to a class centroid is proportional to the expected distance to samples drawn from that class (Eq. 4)
    For an isotropic Gaussian, E||z - z'||² = ||z - μ||² + qσ²; the additive qσ² is dropped. Section III-A.
  • ad hoc to paper The surrogate ψ(x) (uniform resampling in [-b, b], probability p_resample) draws an even number of samples from each class, benign and malicious
    Section III-A: 'It assumed that an even number of samples of each class...'; this invented balance is what converts Eq. 4 into the pairwise loss of Eq. 5.
  • domain assumption All training data is benign
    Definition of D_train in Section III-A; the authors state this as a limitation in Section VI.
  • domain assumption The majority of samples in deployment belong to a single benign class distribution
    Section IV: 'this approach is only viable under the assumption that the majority of samples belong to a single class distribution (benign traffic)'.
  • ad hoc to paper With cosine distance the loss optimizes a von Mises-Fisher analogue of the Gaussian objective
    Section III-A: 'It can be shown that using the cosine distance metric optimises a similar objective...'; asserted without proof.
  • domain assumption Batch Monte Carlo statistics estimate the population expectation in Eq. 4
    Eqs. 4-5 replace expectations with batch averages; finite-sample bias is not discussed.
invented entities (1)
  • Surrogate malicious distribution ψ(x) (uniform feature resampling of benign flows) no independent evidence
    purpose: Serves as the negative class during pretraining so the hinge loss pushes non-benign regions away from the benign cluster (Section III-A).
    No external evidence that resampled benign features resemble real attack traffic; the only support is the AUROC reported on Lycos2017, so the entity has no falsifiable handle outside this paper.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Contrastive Self-Supervised Network Intrusion Detection using Augmented Negative Pairs." pith.science (2026). https://pith.science/paper/YB6D74IT

@misc{pith2026250906550,
  author       = {Pith},
  title        = {Pith review of: Contrastive Self-Supervised Network Intrusion Detection using Augmented Negative Pairs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YB6D74IT}},
  note         = {Machine review of arXiv:2509.06550}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Network intrusion detection remains a critical challenge in cybersecurity. While supervised machine learning models achieve state-of-the-art performance, their reliance on large labelled datasets makes them impractical for many real-world applications. Anomaly detection methods, which train exclusively on benign traffic to identify malicious activity, suffer from high false positive rates, limiting their usability. Recently, self-supervised learning techniques have demonstrated improved performance with lower false positive rates by learning discriminative latent representations of benign traffic. In particular, contrastive self-supervised models achieve this by minimizing the distance between similar (positive) views of benign traffic while maximizing it between dissimilar (negative) views. Existing approaches generate positive views through data augmentation and treat other samples as negative. In contrast, this work introduces Contrastive Learning using Augmented Negative pairs (CLAN), a novel paradigm for network intrusion detection where augmented samples are treated as negative views - representing potentially malicious distributions - while other benign samples serve as positive views. This approach enhances both classification accuracy and inference efficiency after pretraining on benign traffic. Experimental evaluation on the Lycos2017 dataset demonstrates that the proposed method surpasses existing self-supervised and anomaly detection techniques in a binary classification task. Furthermore, when fine-tuned on a limited labelled dataset, the proposed approach achieves superior multi-class classification performance compared to existing self-supervised models.

Figures

Figures reproduced from arXiv: 2509.06550 by Christos Tachtatzis, Hanan Hindy, Jack Wilkie, Robert Atkinson.

Figure 1
Figure 1. Figure 1: Overview of the CLAN framework. A neural network is trained on both genuine benign and augmented network traffic to learn the distribution of [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of the latent representations learned by the CLAN loss [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

40 extracted references · 30 canonical work pages · 2 internal anchors

  1. [1]

    A taxonomy of network threats and the effect of current datasets on intrusion detection systems,

    H. Hindy, D. Brosset, E. Bayne, A. K. Seeam, C. Tachtatzis, R. Atkinson, and X. Bellekens, “A taxonomy of network threats and the effect of current datasets on intrusion detection systems,” IEEE Access, vol. 8, p. 104650–104675, 2020. [Online]. Available: http://dx.doi.org/10.1109/ACCESS.2020.3000179

  2. [2]

    Explainable cross-domain evaluation of ml-based network intrusion detection systems,

    S. Layeghy and M. Portmann, “Explainable cross-domain evaluation of ml-based network intrusion detection systems,”Computers and Electrical Engineering, vol. 108, p. 108692, May 2023. [Online]. Available: http://dx.doi.org/10.1016/j.compeleceng.2023.108692

  3. [3]

    Towards an effective zero-day attack detection using outlier-based deep learning techniques,

    H. Hindy, R. Atkinson, C. Tachtatzis, J.-N. Colin, E. Bayne, and X. Bellekens, “Towards an effective zero-day attack detection using outlier-based deep learning techniques,” 06 2020

  4. [4]

    Anomaly detection using replicator neural networks trained on examples of one class,

    H. A. Dau, V . Ciesielski, and A. Song, “Anomaly detection using replicator neural networks trained on examples of one class,” inSimulated Evolution and Learning, G. Dick, W. N. Browne, P. Whigham, M. Zhang, L. T. Bui, H. Ishibuchi, Y . Jin, X. Li, Y . Shi, P. Singh, K. C. Tan, and K. Tang, Eds. Cham: Springer International Publishing, 2014, pp. 311–322

  5. [5]

    A cookbook of self-supervised learning,

    R. Balestriero, M. Ibrahim, V . Sobal, A. Morcos, S. Shekhar, T. Goldstein, F. Bordes, A. Bardes, G. Mialon, Y . Tian, A. Schwarzschild, A. G. Wilson, J. Geiping, Q. Garrido, P. Fernandez, A. Bar, H. Pirsiavash, Y . LeCun, and M. Goldblum, “A cookbook of self-supervised learning,”

  6. [6]

    Conflow: Contrast network flow improving class-imbalanced learning in network intrusion detection,

    L. Liu, P. Wang, J. Ruan, and J. Lin, “Conflow: Contrast network flow improving class-imbalanced learning in network intrusion detection,” 04 2022

  7. [7]

    Contrastive learning enhanced intrusion detection,

    Y . Yue, X. Chen, Z. Han, X. Zeng, and Y . Zhu, “Contrastive learning enhanced intrusion detection,”IEEE Transactions on Network and Service Management, vol. 19, no. 4, pp. 4232–4247, 2022

  8. [8]

    Sscl-ids: Enhancing generalization of intrusion detection with self-supervised contrastive learning,

    P. Golchin, N. Rafiee, M. Hajizadeh, A. Khalil, R. Kundel, and R. Steinmetz, “Sscl-ids: Enhancing generalization of intrusion detection with self-supervised contrastive learning,” in2024 IFIP Networking Conference (IFIP Networking), 2024, pp. 404–412

  9. [9]

    Network intrusion detection model based on improved byol self-supervised learning,

    Z. Wang, Z. Li, J. Wang, and D. Li, “Network intrusion detection model based on improved byol self-supervised learning,”Security and Communication Networks, vol. 2021, no. 1, p. 9486949, 2021. [Online]. Available: https://onlinelibrary.wiley.com/doi/abs/10.1155/2021/9486949

  10. [10]

    An investigation into the performance of non-contrastive self-supervised learning methods for network intrusion detection,

    H. Fard, T. Schalau, and G. Wunder, “An investigation into the performance of non-contrastive self-supervised learning methods for network intrusion detection,” EasyChair Preprint 14457, EasyChair, 2024

  11. [11]

    An intrusion detection model based on feature reduction and convolutional neural networks,

    Y . Xiao, C. Xing, T. Zhang, and Z. Zhao, “An intrusion detection model based on feature reduction and convolutional neural networks,”IEEE Access, vol. 7, pp. 42 210–42 219, 2019

  12. [12]

    A cnn-lstm model for intrusion detection system from high dimensional data,

    P. Kottapalle, “A cnn-lstm model for intrusion detection system from high dimensional data,”Journal of Information and Computational Science, vol. 10, pp. 1362–1370, 03 2020

  13. [13]

    Hassen and P

    M. Hassen and P. K. Chan,Learning a Neural-network-based Representation for Open Set Recognition, pp. 154–162. [Online]. Available: https://epubs.siam.org/doi/abs/10.1137/1.9781611976236.18

  14. [14]

    A grassmannian approach to zero-shot learning for network intrusion detection,

    J. Rivero, B. Ribeiro, N. Chen, and F. S. Leite, “A grassmannian approach to zero-shot learning for network intrusion detection,” inNeural Information Processing, D. Liu, S. Xie, Y . Li, D. Zhao, and E.-S. M. El- Alfy, Eds. Cham: Springer International Publishing, 2017, pp. 565–575

  15. [15]

    Anomaly based unknown intrusion detection in endpoint environments,

    S. Kim, C. Hwang, and T. Lee, “Anomaly based unknown intrusion detection in endpoint environments,”Electronics, vol. 9, no. 6, 2020. [Online]. Available: https://www.mdpi.com/2079-9292/9/6/1022

  16. [16]

    Network intrusion detector based on isolation . . . forest algorithm,

    S. S, S. G, and B. Priya, “Network intrusion detector based on isolation . . . forest algorithm,” in2022 1st International Conference on Computational Science and Technology (ICCST), 2022, pp. 932–935

  17. [17]

    Unknown attack detection based on zero-shot learning,

    Z. Zhang, Q. Liu, S. Qiu, S. Zhou, and C. Zhang, “Unknown attack detection based on zero-shot learning,”IEEE Access, vol. 8, pp. 193 981– 193 991, 2020

  18. [18]

    Deep unsupervised anomaly detec- tion,

    T. Li, Z. Wang, S. Liu, and W.-Y . Lin, “Deep unsupervised anomaly detec- tion,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE, 2021, pp. 3925–3934. [Online]. Available: https://openaccess.thecvf.com/content/WACV2021/html/Li_ Deep_Unsupervised_Anomaly_Detection_W ACV_2021_paper.html

  19. [19]

    Deep Learning for Network Anomaly Detection under Data Contamination: Evaluating Robustness and Mitigating Performance Degradation

    D. K. Nkashama, J. M. Félicien, A. Soltani, J.-C. Verdier, P.-M. Tardif, M. Frappier, and F. Kabanza, “Deep learning for network anomaly detection under data contamination: Evaluating robustness and mitigating performance degradation,” 2024. [Online]. Available: https://arxiv.org/abs/2407.08838

  20. [20]

    Deep learning approach combining sparse autoencoder with svm for network intrusion detection,

    M. Al-Qatf, Y . Lasheng, M. Al-Habib, and K. Al-Sabahi, “Deep learning approach combining sparse autoencoder with svm for network intrusion detection,”IEEE Access, vol. 6, pp. 52 843–52 856, 2018

  21. [21]

    Deep one-class classification,

    L. Ruff, R. Vandermeulen, N. Goernitz, L. Deecke, S. A. Siddiqui, A. Binder, E. Müller, and M. Kloft, “Deep one-class classification,” in Proceedings of the 35th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, J. Dy and A. Krause, Eds., vol. 80. PMLR, 10–15 Jul 2018, pp. 4393–4402. [Online]. Available: https://...

  22. [22]

    Deep Autoencoding GMM-based Unsupervised Anomaly Detection in Acoustic Signals and its Hyper-parameter Optimization

    H. Purohit, R. Tanabe, T. Endo, K. Suefusa, Y . Nikaido, and Y . Kawaguchi, “Deep autoencoding gmm-based unsupervised anomaly detection in acoustic signals and its hyper-parameter optimization,” 2020. [Online]. Available: https://arxiv.org/abs/2009.12042

  23. [23]

    Efficient malware originated traffic classification by using generative adversarial networks,

    Z. Liu, S. Li, Y . Zhang, X. Yun, and Z. Cheng, “Efficient malware originated traffic classification by using generative adversarial networks,” in2020 IEEE Symposium on Computers and Communications (ISCC), 2020, pp. 1–7

  24. [24]

    Network intrusion detection based on supervised adversarial variational auto-encoder with regularization,

    Y . Yang, K. Zheng, B. Wu, Y . Yang, and X. Wang, “Network intrusion detection based on supervised adversarial variational auto-encoder with regularization,”IEEE Access, vol. 8, pp. 42 169–42 184, 2020

  25. [25]

    Llama: Open and efficient foundation language models,

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozière, N. Goyal, E. Hambro, F. Azhar, A. Rodriguez, A. Joulin, E. Grave, and G. Lample, “Llama: Open and efficient foundation language models,” 2023. [Online]. Available: https://arxiv.org/abs/2302.13971

  26. [26]

    Colorful image colorization,

    R. Zhang, P. Isola, and A. A. Efros, “Colorful image colorization,” CoRR, vol. abs/1603.08511, 2016. [Online]. Available: http://arxiv.org/ abs/1603.08511

  27. [27]

    A simple framework for contrastive learning of visual representations,

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple framework for contrastive learning of visual representations,” 2020. [Online]. Available: https://arxiv.org/abs/2002.05709

  28. [28]

    Infonce: Identifying the gap between theory and practice,

    E. Rusak, P. Reizinger, A. Juhos, O. Bringmann, R. S. Zimmermann, and W. Brendel, “Infonce: Identifying the gap between theory and practice,” 2024. [Online]. Available: https://arxiv.org/abs/2407.00143

  29. [29]

    Bootstrap your own latent: A new approach to self-supervised learning,

    J.-B. Grill, F. Strub, F. Altché, C. Tallec, P. H. Richemond, E. Buchatskaya, C. Doersch, B. A. Pires, Z. D. Guo, M. G. Azar, B. Piot, K. Kavukcuoglu, R. Munos, and M. Valko, “Bootstrap your own latent: A new approach to self-supervised learning,” 2020. [Online]. Available: https://arxiv.org/abs/2006.07733

  30. [30]

    Exploring simple siamese representation learning,

    X. Chen and K. He, “Exploring simple siamese representation learning,”

  31. [31]

    Vicreg: Variance-invariance- covariance regularization for self-supervised learning,

    A. Bardes, J. Ponce, and Y . LeCun, “Vicreg: Variance-invariance- covariance regularization for self-supervised learning,” 2022. [Online]. Available: https://arxiv.org/abs/2105.04906

  32. [32]

    Barlow twins: Self-supervised learning via redundancy reduction,

    J. Zbontar, L. Jing, I. Misra, Y . LeCun, and S. Deny, “Barlow twins: Self-supervised learning via redundancy reduction,” 2021. [Online]. Available: https://arxiv.org/abs/2103.03230

  33. [34]

    Dimensionality reduction by learning an invariant mapping,

    R. Hadsell, S. Chopra, and Y . LeCun, “Dimensionality reduction by learning an invariant mapping,” in2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, ser. CVPR’06, vol. 2, 2006, pp. 1735–1742

  34. [35]

    Learning a similarity metric discriminatively, with application to face verification,

    S. Chopra, R. Hadsell, and Y . LeCun, “Learning a similarity metric discriminatively, with application to face verification,” in2005 IEEE Computer Society Conference on Computer Vision and Pattern Recogni- tion, ser. CVPR’05, vol. 1, 2005, pp. 539–546 vol. 1

  35. [36]

    Improved deep metric learning with multi-class n-pair loss objective,

    K. Sohn, “Improved deep metric learning with multi-class n-pair loss objective,” inAdvances in Neural Information Processing Systems, D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett, Eds., vol. 29. Curran Associates, Inc., 2016. [Online]. Available: https://proceedings.neurips.cc/paper_files/paper/ 2016/file/6b180037abbebea991d8b1232f8a8ca9-Paper.pdf

  36. [37]

    From cic-ids2017 to lycos-ids2017: A corrected dataset for better performance,

    A. ROSAY , F. CARLIER, E. CHEV AL, and P. LEROUX, “From cic-ids2017 to lycos-ids2017: A corrected dataset for better performance,” inIEEE/WIC/ACM International Conference on Web Intelligence and Intelligent Agent Technology, ser. WI-IAT ’21. New York, NY , USA: Association for Computing Machinery, 2022, p. 570–575. [Online]. Available: https://doi.org/10....

  37. [38]

    Toward generating a new intrusion detection dataset and intrusion traffic characterization,

    I. Sharafaldin., A. Habibi Lashkari., and A. A. Ghorbani., “Toward generating a new intrusion detection dataset and intrusion traffic characterization,” inProceedings of the 4th International Conference on Information Systems Security and Privacy - ICISSP ,, INSTICC. SciTePress, 2018, pp. 108–116

  38. [2020]

    Available: https://arxiv.org/abs/2011.10566

    [Online]. Available: https://arxiv.org/abs/2011.10566

  39. [2021]

    Available: https://arxiv.org/abs/2102.06810

    [Online]. Available: https://arxiv.org/abs/2102.06810

  40. [2023]

    Available: https://arxiv.org/abs/2304.12210

    [Online]. Available: https://arxiv.org/abs/2304.12210

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.