Pith. sign in

REVIEW 2 major objections 6 minor 46 references

Addressing malware family concept drift with triplet autoencoder

T0 review · 2 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper claims that a triplet-loss autoencoder, clustered with DBSCAN in its latent space, can flag malware samples that belong to no known family, reaching overall F1 0.81 on the Drebin Android dataset and beating two drift-detection…

desk verdict A coherent open-set malware family detector whose evaluation never tests temporal drift: the held-out family is withheld by label, not by timestamp, so the central claim does not follow. read the letter →

arxiv 2507.00348 v1 pith:PIPQF3JJ submitted 2025-07-01 cs.CR

classification cs.CR
keywords conceptdriftmalwarefamilydetectiontripletlossautoencoderDBSCANopen-setrecognitionAndroidWindowsPE
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

The paper is trying to establish that a malware detector can tell when a sample belongs to no known family, which is the core of coping with concept drift caused by new malware families. The proposed method compresses each sample with a triplet-trained autoencoder so that same-family samples form tight clusters and different families are pushed apart, then runs DBSCAN inside that latent space to find sub-clusters and set per-family distance boundaries. At test time, a sample whose distance to the nearest known centroid exceeds the boundary is flagged as a possible new family rather than being mislabeled. On the Drebin Android dataset the method reaches an overall F1 of 0.81 for withheld families, against 0.78 for the same embedding with a statistically derived threshold and 0.62 for the CADE contrastive baseline; on the Windows PE dataset BODMAS it reports per-family F1 scores from 0.51 to 0.99. The significance is that drift can be handled without frequent retraining and without manually tuning a rejection threshold.

What carries the argument

The central object is a triplet autoencoder: an autoencoder whose 32-dimensional bottleneck is trained with both reconstruction loss and triplet loss $L = \max\{0, D(x_a, x_p) - D(x_a, x_n) + \alpha\}$, so that the latent space is both low-dimensional and metric-structured. On top of this, DBSCAN is the load-bearing mechanism: it discovers multiple clusters within a single malware family (for example, two clusters in FakeInstaller, reducing the mean sample-to-centroid distance by 26 percent), marks outliers, and produces centroids and per-family thresholds defined as the maximum distance from a centroid to any core point. The threshold is what converts the embedding into an open-set decision rule: within threshold means known family, beyond means likely new family.

What would settle it

Run the same Drebin protocol but restrict each withheld family's test set to samples created after the training cutoff; if the DBSCAN method's overall F1 no longer beats the MAD and CADE baselines, the claim that it detects temporally emerging families would be contradicted.

Watch

Extended reading notes

Core claim

The central claim is that the appearance of new malware families, one of the two named forms of concept drift, can be detected by learning a metric space instead of a fixed classifier. A triplet autoencoder maps malware into a 32-dimensional embedding in which known families form compact, separated clusters; DBSCAN then exposes sub-clusters inside any single family, discards outliers, and defines each cluster's centroid and membership threshold as the maximum distance from the centroid to a core point. A test sample is assigned to the nearest known family only if it falls within that learned boundary; otherwise it is labeled unknown and sent for expert review. The paper reports that this DBSCAN-based threshold consistently beats the MAD threshold on the same embedding, with overall F1 of 0.81 versus 0.78 on Drebin and 0.62 for the CADE baseline, and that the approach transfers to the Windows PE dataset BODMAS with per-family F1 between 0.51 and 0.99. The authors present the automatic threshold as the practical advantage: no empirical coefficient is needed.

Load-bearing premise

The evaluation rests on treating a held-out malware family as genuinely unseen at test time even though its samples' creation timestamps overlap the training period, because no temporal split is applied within the withheld family.

Editorial extensions

If this is right

  • A deployed detector can flag malware as belonging to no known family instead of forcing a wrong family label, allowing new samples to be routed to expert analysis.
  • Because DBSCAN sets each family's threshold automatically, the method removes the empirical coefficient that a statistical dispersion threshold such as MAD requires.
  • The same triplet-autoencoder-plus-DBSCAN pipeline reports strong F1 scores on both Android and Windows PE malware, indicating the approach transfers across feature representations and operating systems.
  • Sub-clustering inside a family fixes a known failure of single-centroid methods: for FakeInstaller, DBSCAN found two clusters and reduced the mean sample-to-centroid distance by 26 percent, so distances and thresholds become more trustworthy.
  • The method targets the emergence of new families rather than gradual evolution within existing families, so it complements rather than replaces drift detection aimed at shifting distributions.

Reading between the lines

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

  • The paper does not apply a temporal split inside the withheld family; a stricter evaluation would train on each family's early samples and test only on that family's later samples, and the reported F1 advantage may shrink under that protocol.
  • The threshold is fixed at the maximum core-point distance, so the boundary is sensitive to the most extreme inlier in each cluster; testing a quantile-based boundary could improve the worst BODMAS families (mira at 0.58 and sfone at 0.51).
  • Only one family is withheld at a time; when several new families appear together, DBSCAN's noise points could be clustered to recognize multiple simultaneous unknowns, an extension the paper leaves for future work.
  • Flagged unknowns naturally feed an active-learning loop: an analyst labels rejected samples and the centroids are recomputed, so the system can absorb a new family without a full retrain.
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

2 major / 6 minor

Summary. The paper proposes a triplet-autoencoder plus DBSCAN method for detecting new malware families, framed as addressing concept drift caused by the emergence of previously unseen families. The method trains an autoencoder with a combination of reconstruction loss and triplet loss, clusters the latent space with DBSCAN, computes per-cluster centroids and distance thresholds, and labels test samples as known or unknown based on proximity to the nearest centroid. The authors evaluate on the Drebin Android dataset and the BODMAS Windows PE dataset, comparing against CADE and a MAD-threshold variant, and report higher F1 scores for their DBSCAN-based approach.

Significance. If the evaluation were sound, the paper would offer a useful and relatively simple contribution to malware family detection under drift, with the strength of explicit sub-cluster modeling via DBSCAN and evaluation on two datasets. The method is clearly described, and the comparison with CADE and MAD is reasonable. However, the central claimed contribution is temporal concept drift, and the reported experimental protocol does not actually isolate temporal emergence of new families. The significance of the empirical results is therefore not established as stated.

major comments (2)
  1. [Section IV-C, Tables III and IV] The drift simulation uses the entire held-out malware family as the unknown set regardless of timestamps, while the temporal 80:20 split is applied only to the remaining families. For example, in Table III the FakeInstaller row uses all 925 FakeInstaller samples as unknown, even though a large fraction of those samples fall within the training time window; the known set consists of the time-based test slice of the other families. This protocol measures arbitrary family withholding (leave-one-family-out open-set recognition), not the temporal emergence of a new family after the training cutoff, which is the paper's stated central claim. The experiments should be re-run with the unknown set restricted to held-out-family samples occurring after the training cutoff, or the paper should be explicitly reframed as addressing open-set family detection rather than concept drift.
  2. [Table IV] The reported 'No. of known samples' values are not consistent with the stated per-family 80:20 temporal split and the family sizes in Table II. For the small row, the remaining families contain 12,852 samples, so a 20% test split would give approximately 2,570 known samples, not the reported 1,473; for the berbew row, the expected value is approximately 2,943, not 2,817. If the authors instead use a global date-based split, the exact cutoff dates and per-family test counts must be reported. As written, Table IV cannot be reproduced from the described experimental protocol.
minor comments (6)
  1. [Tables I and III] The family name is written as 'GingerMaster' in Table I but 'GinMaster' in Table III; please use a single consistent name.
  2. [Equations (1) and (2)] Equation (2) defines D as squared Euclidean distance, but the text and the threshold discussion refer to plain distances between samples and centroids; please clarify which distance is actually used in the centroid and threshold computations.
  3. [Section V-A] The MAD baseline is said to require an empirically determined coefficient, but the coefficient value is never reported; please include it for reproducibility.
  4. [Section III, 'Impact of DBSCAN'] The claim that DBSCAN 'calculates the threshold automatically without requiring a manual coefficient' should be qualified, since the DBSCAN parameters epsilon and minPts are still chosen by heuristics (the k-distance plot and the 2 x dim rule).
  5. [Section IV-B] The paper states that the BODMAS training and test sets are split by first-seen timestamps from August 2019 to September 2020, but Table II does not report how many samples of each family fall into the test window; adding per-family test counts would make the setup clearer.
  6. [References] Reference [2] is spelled 'A V-Test' but should be 'AV-TEST'; please correct the citation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the F1 evaluation is on held-out families with thresholds derived from training data alone; the temporal-split weakness is an experimental-validity concern, not a circular derivation.

full rationale

The derivation chain is self-contained. The model is trained on known families with one family excluded, and DBSCAN centroids and per-family thresholds are computed exclusively from the training latent space: "Each family's threshold is determined by the distance from the centroid to the furthest point within the cluster." Test samples are then classified by comparing their distance to the nearest centroid against that threshold. The reported F1 scores in Tables III and IV are computed on samples whose family labels were never used to set the thresholds, so the "new family detection" result is not equivalent to a fitted parameter or to the training labels. The only self-citation, reference [25], is a metric-learning BCI paper co-authored by one of the present authors and appears only in related work as an example application; it is not load-bearing. Section VI openly acknowledges limitations such as considering only one unknown family and the possibility of fine-tuning DBSCAN's epsilon and minPts; these are honest scope statements rather than circular moves. The skeptic's concern that the held-out family's early samples are contemporaneous with the training window, based on the statement in Section IV that "we simulate the presence of drifting samples by systematically excluding one malware family from the training data," is a threat to the temporal-drift interpretation of the experiment, but it does not make the predicted F1 scores an input of the method. The apparent BODMAS row inconsistency in Table IV (for 'small', 1473 known + 3606 unknown = 5079, while Table II reports 3606 samples for that family) is a reporting or preprocessing consistency issue, not a circular derivation. No step in the paper reduces by construction to its own inputs.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The method introduces no new entities. It relies on a handful of domain assumptions about how malware families manifest in a static-feature latent space, plus several unreported hyperparameters (epsilon, minPts, margin, loss weighting, variance threshold) that the reported F1 scores depend on.

free parameters (6)
  • DBSCAN epsilon = not reported
    Chosen via k-distance plot on training latent space; controls cluster membership and threshold, directly affects new-family detection.
  • DBSCAN minPts = 2 x bottleneck_dim = 64 (for 32-d latent)
    Set by heuristic as twice the latent dimensionality; determines core points.
  • Triplet loss margin alpha = not reported
    Margin in Eq. (1); affects embedding separability.
  • Reconstruction vs triplet loss weighting = not reported
    Autoencoder is trained with a combination of losses; no weights given.
  • Feature variance threshold = not reported
    Variance filtering removes low-variance features; threshold not specified.
  • MAD coefficient for baseline = not reported
    Comparison baseline requires an empirical coefficient; value used is not disclosed.
assumptions (4)
  • domain assumption All inputs are malware samples; only family labels matter.
    Explicitly stated in Section I; excludes benign/malicious binary distinction.
  • domain assumption Emergence of a new family manifests as out-of-distribution samples in a 32-d metric space learned from static features.
    Core premise that distances to cluster centroids separate known from unknown families.
  • domain assumption DBSCAN clusters in the latent space correspond to meaningful subfamilies or variants.
    Used to compute centroids and thresholds; false clusters would distort boundaries.
  • domain assumption Static features (Drebin manifest/code features, BODMAS feature vector) are sufficient to detect family drift.
    No dynamic analysis; packed samples excluded because static features are unreliable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Addressing malware family concept drift with triplet autoencoder." pith.science (2026). https://pith.science/paper/PIPQF3JJ

@misc{pith2026250700348,
  author       = {Pith},
  title        = {Pith review of: Addressing malware family concept drift with triplet autoencoder},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PIPQF3JJ}},
  note         = {Machine review of arXiv:2507.00348}
}
read the original abstract

Machine learning is increasingly vital in cybersecurity, especially in malware detection. However, concept drift, where the characteristics of malware change over time, poses a challenge for maintaining the efficacy of these detection systems. Concept drift can occur in two forms: the emergence of entirely new malware families and the evolution of existing ones. This paper proposes an innovative method to address the former, focusing on effectively identifying new malware families. Our approach leverages a supervised autoencoder combined with triplet loss to differentiate between known and new malware families. We create clear and robust clusters that enhance the accuracy and resilience of malware family classification by utilizing this metric learning technique and the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm. The effectiveness of our method is validated using an Android malware dataset and a Windows portable executable (PE) malware dataset, showcasing its capability to sustain model performance within the dynamic landscape of emerging malware threats. Our results demonstrate a significant improvement in detecting new malware families, offering a reliable solution for ongoing cybersecurity challenges.

Figures

Figures reproduced from arXiv: 2507.00348 by the authors.

Figure 1
Figure 1. An overview of the method. We employ metric learning, a machine learning approach that focuses on defining a distance metric between data points to transform the data space. This method aims to bring similar points closer together while pushing dissimilar points further apart [40]. In our implementation, we utilize triplet loss, which simultaneously considers pairs of similar and dissimilar points, enhancing the mod… view at source ↗
Figure 2
Figure 2. Clustering results using DBSCAN algorithm. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Boxplot diagrams showing the distances between samples and their family centroids for three feature representations: original features, vanilla [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: These t-SNE visualizations offer a clearer picture of [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 4
Figure 4. Figure 4: t-SNE diagrams of original features space, vanilla autoencoder and triplet autoencoder. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 41 canonical work pages

  1. [1]

    Learning under concept drift: A review,

    J. Lu, A. Liu, F. Dong, F. Gu, J. Gama, and G. Zhang, “Learning under concept drift: A review,” IEEE transactions on knowledge and data engineering, vol. 31, no. 12, pp. 2346– 2363, 2018

  2. [2]

    Malware statistics & trends report,

    A V-Test, “Malware statistics & trends report,” Accessed: 14 July 2024, 2023, [Online]. Available: https://www.av-test.org/ en/statistics/malware/

  3. [3]

    Tesseract: Eliminating experimental bias in malware classification across space and time,

    F. Pendlebury, F. Pierazzi, R. Jordaney, J. Kinder, and L. Cav- allaro, “Tesseract: Eliminating experimental bias in malware classification across space and time,” in 28th USENIX security symposium (USENIX Security 19) , 2019, pp. 729–746

  4. [4]

    Transcend: Detecting concept drift in malware classification models,

    R. Jordaney et al. , “Transcend: Detecting concept drift in malware classification models,” in 26th USENIX security sym- posium (USENIX security 17) , 2017, pp. 625–642

  5. [5]

    Towards open set deep networks,

    A. Bendale and T. E. Boult, “Towards open set deep networks,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 1563–1572

  6. [6]

    A simple unified frame- work for detecting out-of-distribution samples and adversarial attacks,

    K. Lee, K. Lee, H. Lee, and J. Shin, “A simple unified frame- work for detecting out-of-distribution samples and adversarial attacks,” Advances in neural information processing systems , vol. 31, 2018

  7. [7]

    Android malware detection: Mission accomplished? a review of open challenges and future per- spectives,

    A. Guerra-Manzanares, “Android malware detection: Mission accomplished? a review of open challenges and future per- spectives,” Computers & Security , p. 103 654, 2023

  8. [8]

    Novel feature extraction, selection and fusion for effective malware family classification,

    M. Ahmadi, D. Ulyanov, S. Semenov, M. Trofimov, and G. Giacinto, “Novel feature extraction, selection and fusion for effective malware family classification,” in Proceedings of the sixth ACM conference on data and application security and privacy, 2016, pp. 183–194

Show all 46 references
  1. [9]

    Malware detection based on mining api calls,

    A. Sami, B. Yadegari, H. Rahimi, N. Peiravian, S. Hashemi, and A. Hamze, “Malware detection based on mining api calls,” in Proceedings of the 2010 ACM symposium on applied computing, 2010, pp. 1020–1025

  2. [10]

    Byte level n–gram analysis for malware detection,

    S. Jain and Y . K. Meena, “Byte level n–gram analysis for malware detection,” in Computer Networks and Intelligent Computing: 5th International Conference on Information Pro- cessing, ICIP 2011, Bangalore, India, August 5-7, 2011. Proceedings, Springer, 2011, pp. 51–59

  3. [11]

    Malware detection and classifica- tion based on n-grams attribute similarity,

    Z. Fuyong and Z. Tiezhu, “Malware detection and classifica- tion based on n-grams attribute similarity,” in2017 IEEE inter- national conference on computational science and engineering (CSE) and IEEE international conference on embedded and ubiquitous computing (EUC), IEEE, vo...

  4. [12]

    Deep android malware detection,

    N. McLaughlin et al. , “Deep android malware detection,” in Proceedings of the seventh ACM on conference on data and application security and privacy , 2017, pp. 301–308

  5. [13]

    Sequential op- code embedding-based malware detection method,

    A. G. Kakisim, S. Gulmez, and I. Sogukpinar, “Sequential op- code embedding-based malware detection method,”Computers & Electrical Engineering , vol. 98, p. 107 703, 2022

  6. [14]

    Malware detection based on deep learn- ing algorithm,

    D. Yuxin and Z. Siyi, “Malware detection based on deep learn- ing algorithm,” Neural Computing and Applications , vol. 31, pp. 461–472, 2019

  7. [15]

    Drebin: Effective and explainable detection of android malware in your pocket.,

    D. Arp, M. Spreitzenbarth, M. Hubner, H. Gascon, K. Rieck, and C. Siemens, “Drebin: Effective and explainable detection of android malware in your pocket.,” in Ndss, vol. 14, 2014, pp. 23–26

  8. [16]

    Ember: An open dataset for training static pe malware machine learning models,

    H. S. Anderson and P. Roth, “Ember: An open dataset for training static pe malware machine learning models,” arXiv preprint arXiv:1804.04637, 2018

  9. [17]

    Bodmas: An open dataset for learning based temporal analysis of pe malware,

    L. Yang, A. Ciptadi, I. Laziuk, A. Ahmadzadeh, and G. Wang, “Bodmas: An open dataset for learning based temporal analysis of pe malware,” in 2021 IEEE Security and Privacy Workshops (SPW), IEEE, 2021, pp. 78–84

  10. [18]

    Maar: Robust features to detect malicious activity based on api calls, their arguments and return values,

    Z. Salehi, A. Sami, and M. Ghiasi, “Maar: Robust features to detect malicious activity based on api calls, their arguments and return values,” Engineering Applications of Artificial In- telligence, vol. 59, pp. 93–102, 2017

  11. [19]

    Malware detection and classification based on extraction of api sequences,

    D. Uppal, R. Sinha, V . Mehra, and V . Jain, “Malware detection and classification based on extraction of api sequences,” in 2014 International conference on advances in computing, com- munications and informatics (ICACCI), IEEE, 2014, pp. 2337– 2342

  12. [20]

    Detecting obfus- cated malware using reduced opcode set and optimised runtime trace,

    P. O’kane, S. Sezer, and K. McLaughlin, “Detecting obfus- cated malware using reduced opcode set and optimised runtime trace,” Security Informatics, vol. 5, pp. 1–12, 2016

  13. [21]

    Network malware classification comparison using dpi and flow packet headers,

    A. Boukhtouta, S. A. Mokhov, N.-E. Lakhdari, M. Debbabi, and J. Paquet, “Network malware classification comparison using dpi and flow packet headers,” Journal of Computer Virology and Hacking Techniques, vol. 12, pp. 69–100, 2016

  14. [22]

    Malicious software classification using transfer learning of resnet-50 deep neural network,

    E. Rezende, G. Ruppert, T. Carvalho, F. Ramos, and P. De Geus, “Malicious software classification using transfer learning of resnet-50 deep neural network,” in 2017 16th IEEE international conference on machine learning and applications (ICMLA), IEEE, 2017, pp. 1011–1014

  15. [23]

    In defense of the triplet loss for person re-identification,

    A. Hermans, L. Beyer, and B. Leibe, “In defense of the triplet loss for person re-identification,” arXiv preprint arXiv:1703.07737, 2017

  16. [24]

    Triplet loss in siamese network for object tracking,

    X. Dong and J. Shen, “Triplet loss in siamese network for object tracking,” in Proceedings of the European conference on computer vision (ECCV) , 2018, pp. 459–474

  17. [25]

    A zero-shot deep metric learning approach to brain–computer interfaces for image retrieval,

    B. McCartney, B. Devereux, and J. Martinez-del-Rincon, “A zero-shot deep metric learning approach to brain–computer interfaces for image retrieval,” Knowledge-Based Systems , vol. 246, p. 108 556, 2022

  18. [26]

    Representation learning with contrastive predictive coding,

    A. v. d. Oord, Y . Li, and O. Vinyals, “Representation learning with contrastive predictive coding,” arXiv preprint arXiv:1807.03748, 2018

  19. [27]

    Metric learning- based multimodal audio-visual emotion recognition,

    E. Ghaleb, M. Popa, and S. Asteriadis, “Metric learning- based multimodal audio-visual emotion recognition,” IEEE Multimedia, vol. 27, no. 1, pp. 37–48, 2019

  20. [28]

    In defence of metric learning for speaker recognition,

    J. S. Chung et al., “In defence of metric learning for speaker recognition,” arXiv preprint arXiv:2003.11982 , 2020

  21. [29]

    Multi-instance multi- label distance metric learning for genome-wide protein func- tion prediction,

    Y . Xu, H. Min, H. Song, and Q. Wu, “Multi-instance multi- label distance metric learning for genome-wide protein func- tion prediction,” Computational biology and chemistry, vol. 63, pp. 30–40, 2016

  22. [30]

    A novel drug repositioning approach based on collaborative metric learning,

    H. Luo, J. Wang, C. Yan, M. Li, F.-X. Wu, and Y . Pan, “A novel drug repositioning approach based on collaborative metric learning,” IEEE/ACM transactions on computational biology and bioinformatics, vol. 18, no. 2, pp. 463–471, 2019

  23. [31]

    Contrastive learning for robust android malware familial classification,

    Y . Wu, S. Dou, D. Zou, W. Yang, W. Qiang, and H. Jin, “Contrastive learning for robust android malware familial classification,” IEEE Transactions on Dependable and Secure Computing, 2022

  24. [32]

    Application of distance metric learning to automated malware detection,

    M. Jure ˇcek and R. Lórencz, “Application of distance metric learning to automated malware detection,”IEEE Access, vol. 9, pp. 96 151–96 165, 2021

  25. [33]

    Fewm- hgcl: Few-shot malware variants detection via heterogeneous graph contrastive learning,

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

  26. [34]

    Autoencoder- based deep metric learning for network intrusion detection,

    G. Andresini, A. Appice, and D. Malerba, “Autoencoder- based deep metric learning for network intrusion detection,” Information Sciences, vol. 569, pp. 706–727, 2021

  27. [35]

    Tracking concept drift in malware families,

    A. Singh, A. Walenstein, and A. Lakhotia, “Tracking concept drift in malware families,” in Proceedings of the 5th ACM workshop on Security and artificial intelligence, 2012, pp. 81– 92

  28. [36]

    Transcending transcend: Revisiting malware classification in the presence of concept drift,

    F. Barbero, F. Pendlebury, F. Pierazzi, and L. Cavallaro, “Transcending transcend: Revisiting malware classification in the presence of concept drift,” in 2022 IEEE Symposium on Security and Privacy (SP) , IEEE, 2022, pp. 805–823

  29. [37]

    Cade: Detecting and explaining concept drift samples for security applications,

    L. Yang et al., “Cade: Detecting and explaining concept drift samples for security applications,” in 30th USENIX Security Symposium (USENIX Security 21) , 2021, pp. 2327–2344

  30. [38]

    Insomnia: Towards concept-drift robustness in network intrusion detection,

    G. Andresini, F. Pendlebury, F. Pierazzi, C. Loglisci, A. Appice, and L. Cavallaro, “Insomnia: Towards concept-drift robustness in network intrusion detection,” in Proceedings of the 14th ACM workshop on artificial intelligence and security , 2021, pp. 111–122

  31. [39]

    Temporal analysis of dis- tribution shifts in malware classification for digital forensics,

    F. Zola, J. L. Bruse, and M. Galar, “Temporal analysis of dis- tribution shifts in malware classification for digital forensics,” in 2023 IEEE European Symposium on Security and Privacy Workshops (EuroS&PW), IEEE, 2023, pp. 439–450

  32. [40]

    Deep metric learning: A survey,

    M. Kaya and H. ¸ S. Bilge, “Deep metric learning: A survey,” Symmetry, vol. 11, no. 9, p. 1066, 2019

  33. [41]

    The curse (s) of dimension- ality,

    N. Altman and M. Krzywinski, “The curse (s) of dimension- ality,” Nat Methods, vol. 15, no. 6, pp. 399–400, 2018

  34. [42]

    Dbscan revisited, revisited: Why and how you should (still) use dbscan,

    E. Schubert, J. Sander, M. Ester, H. P. Kriegel, and X. Xu, “Dbscan revisited, revisited: Why and how you should (still) use dbscan,” ACM Transactions on Database Systems (TODS), vol. 42, no. 3, pp. 1–21, 2017

  35. [43]

    The k-means algo- rithm: A comprehensive survey and performance evaluation,

    M. Ahmed, R. Seraj, and S. M. S. Islam, “The k-means algo- rithm: A comprehensive survey and performance evaluation,” Electronics, vol. 9, no. 8, p. 1295, 2020

  36. [44]

    A density-based algorithm for discovering clusters in large spatial databases with noise,

    M. Ester et al. , “A density-based algorithm for discovering clusters in large spatial databases with noise,” in KDD, vol. 96, 1996, pp. 226–231

  37. [45]

    Fesa: Feature selection architecture for ransomware detection under concept drift,

    D. W. Fernando and N. Komninos, “Fesa: Feature selection architecture for ransomware detection under concept drift,” Computers & Security , vol. 116, p. 102 659, 2022

  38. [46]

    Visualizing data using t- sne.,

    L. Van der Maaten and G. Hinton, “Visualizing data using t- sne.,” Journal of machine learning research , vol. 9, no. 11, 2008

Pith tools

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