Pith. sign in

REVIEW 3 major objections 7 minor 2 cited by

EMBER2024 -- A Benchmark Dataset for Holistic Evaluation of Malware Classifiers

T0 review · 3 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read EMBER2024 gives malware researchers one benchmark with 3.2M files, six formats, seven label types, and a challenge set of files that initially evaded every antivirus engine.

desk verdict Genuinely useful benchmark artifact that fills a real gap; core claims hold, but a few fixable errors and one substantive concern about cross-week near-duplicate leakage in the challenge set should be addressed before publication. read the letter →

arxiv 2506.05074 v1 pith:KQFLU3F3 submitted 2025-06-05 cs.CR cs.LG

Robert J. Joyce , Gideon Miller , Phil Roth , Richard Zak , Elliott Zaresky-Williams , Hyrum Anderson , Edward Raff , James Holt This is my paper
classification cs.CRcs.LG
keywords malwarebenchmarkdatasetEMBER2024evasivechallengesetmulti-formatclassificationstaticfeatureextractionVirusTotallabelingconceptdriftmachinelearningfor
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

EMBER2024 is a malware-classifier benchmark built from 3,238,315 files first submitted to VirusTotal between September 2023 and December 2024. The paper's central claim is that one public dataset can support holistic evaluation of malware classifiers: six file formats (Win32, Win64, .NET, APK, ELF, and PDF), seven kinds of labels and tags, and a dedicated challenge set of 6,315 files that initially went undetected by every antivirus product yet were later flagged as malicious. This matters because earlier large public benchmarks were mostly single-format, single-task, and ignored the evasive files that make real-world detection hard. The dataset also introduces a new feature format for cross-format representation and releases code to reproduce the collection and labeling pipeline, so results can be compared and the dataset refreshed.

What carries the argument

The load-bearing mechanism is a two-stage VirusTotal labeling protocol. Each file is queried within 24 hours of first submission and again 90 or more days later; files suspected to be benign are re-scanned at least 30 days after submission. Zero detected products at the later scan labels a file benign, five or more independent detections labels it malicious, and files that were undetected at first submission but malicious by the later scan form the challenge set. This temporal delay turns antivirus signature updates into a signal for evasive malware. The other central object is EMBER feature version 3, a 2,568-dimensional vector that extends the prior PE-focused feature format and still extracts general, string, byte-histogram, and byte-entropy features from non-PE or unparsable files, so one representation can cover Win32, Win64, .NET, APK, ELF, and PDF files.

What would settle it

Manually analyze a random sample of files EMBER2024 labels benign by executing them in an isolated sandbox and inspecting network, filesystem, and process behavior; if a nontrivial fraction turn out to be malicious, the benign labels and the challenge-set definition are biased. A cheaper check is to rerun the labeling protocol with an independent second scanner or a stricter detection threshold and see whether reported benchmark rankings on the challenge set change.

Watch

Extended reading notes

Core claim

The authors introduce EMBER2024 as the first public malware benchmark that is simultaneously multi-platform, multi-label, and evasive-aware. Using antivirus verdicts taken within 24 hours of first submission and again 90 or more days later, they label files as benign if no product detects them, malicious if five or more independent products agree, and as a challenge set if the file was undetected at first submission but malicious by re-scan. Baseline gradient-boosted tree classifiers reach roughly 0.995 ROC AUC on the ordinary temporal test set, but their precision-recall scores collapse on the challenge set—for all files combined the PR AUC is 0.5722, and for APK files it falls to 0.2374—which the paper reads as evidence that evasive malware is a measurable, still-open problem. The companion feature version 3 vector, together with released construction code, is what lets this single benchmark cover multiple formats and tasks.

Load-bearing premise

The entire label set—benign, malicious, and challenge—rests on the assumption that antivirus verdicts collected after 30 or more days are reliable ground truth, so any malicious file that no antivirus product ever flags is silently counted as benign.

Editorial extensions

If this is right

  • A researcher can benchmark malware detection, family identification, behavior prediction, packer identification, file-property prediction, exploit tagging, and threat-group attribution on the same train/test split and the same feature format.
  • The challenge set gives a compact standard evaluation for evasive malware; the paper's all-files model achieves PR AUC 0.9971 on the ordinary test set but only 0.5722 on the challenge set.
  • The fixed weekly time split makes concept drift directly testable, including the 75 families that appear in the test set but never in training.
  • Because the feature extractor and dataset-construction pipeline are released, researchers can rerun the methodology on new time windows rather than depending on a static snapshot.
  • Classifiers trained on reduced features for ELF, PDF, and APK still reach high AUC on ordinary test sets, so the multi-format feature approach is viable beyond Windows PE files.

Reading between the lines

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

  • A natural extension the paper does not pursue is per-format challenge sets; the APK challenge PR AUC of 0.2374 suggests evasion mechanisms differ by platform and deserve separate study.
  • The delayed-detection criterion could be graded—files first detected after 30, 90, or 180 days may form an evasion difficulty ladder, letting researchers measure how much longer a classifier resists newer malware.
  • Because only hashes, metadata, and feature vectors are released, not raw file bytes, byte-level deep-learning detectors and adversarial perturbation experiments cannot be run directly on EMBER2024; its holistic scope is mainly for static-feature classifiers.
  • Strong AUC on a temporal split can coexist with weak performance on evasive files, so future benchmark reports that cite EMBER2024 should report challenge-set precision-recall alongside standard AUC.
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

3 major / 7 minor

Summary. The paper presents EMBER2024, a new large-scale benchmark dataset for malware classification, containing 3,238,315 files collected over 64 weeks from six file formats (Win32, Win64, .NET, APK, ELF, PDF). The dataset provides seven types of labels and tags, supports tasks including malware detection, family classification, and behavior prediction, and includes a 'challenge set' of 6,315 malicious files that were initially undetected by all VirusTotal AV engines but later detected. The paper also introduces EMBER feature version 3, releases code for dataset construction and model training, and reports baseline LightGBM results. The central claim is that this is the first holistic benchmark combining multi-format support, multi-task labels, and an evasive-malware evaluation subset, with reproducible construction methodology.

Significance. If the dataset construction is sound, EMBER2024 would be a significant contribution: it addresses the lack of recent, multi-format, multi-task malware benchmarks and is the first to offer a dedicated evasive-malware challenge set. The release of construction code is a strong reproducibility feature, and the baseline results provide a useful reference. However, the paper contains concrete numerical errors and an under-specified aspect of the challenge set's construction that could bias evaluation. These issues are fixable, but they are load-bearing for the paper's reliability.

major comments (3)
  1. [Table 3 and abstract] Table 3 contains an arithmetic error: the PDF row reports a total of 12,805, but its components (52,000 + 12,000 + 805) sum to 64,805. The challenge-set column in Table 3 sums to 6,291, not the stated 6,315, and the file-type totals sum to 3,238,291, not the 3,238,315 given in the abstract and Section 1.1. The paper must correct these numbers and explain the 24-file discrepancy.
  2. [Section 4.1 vs Table 5] Section 4.1 reports a ROC AUC of 0.9949 for the LightGBM classifier trained on the full training set, while Table 5 reports 0.9969 for the 'All files / All files' entry. Since Section 4.2 says all classifiers use the same hyperparameters described in Section 4.1, these numbers should match unless there were separate training runs. Please clarify the source of the discrepancy and, if multiple runs were performed, report the variability.
  3. [Section 2.3] The challenge-set near-duplicate exclusion only covers files from the same week of data collection. Because the training and test sets span 64 weeks, a challenge file may have a TLSH near-duplicate (distance of 30 or less) in a different week's portion of the training set, which would allow a classifier to exploit that near-duplicate rather than generalize to evasive malware. The paper does not report any cross-week TLSH distance analysis between the challenge set and the training/test sets. Please provide this analysis (for example, the distribution of minimum TLSH distances from each challenge file to any train/test file) or justify why cross-week near-duplicates are unlikely. This is essential to support the claim that the challenge set measures robustness against evasive malware.
minor comments (7)
  1. [Section 2.4] The benign label is defined as no AV detection after re-scanning at least 30 days later. This is a noisy ground-truth assumption that also affects the challenge-set definition. The paper should explicitly acknowledge the risk that never-detected malware is labeled benign and that the challenge set depends on the same AV threshold. A short limitations discussion would improve the paper.
  2. [Figure 3] The JSON example is missing a comma after the 'md5' value.
  3. [Section 3.3] The text says the implementation switched from 'LEIF' to pefile; this should be 'LIEF'.
  4. [Table 7] The column header 'weighed averaging' should be 'weighted averaging'.
  5. [References] Reference [20] is incomplete; it lacks the publication venue.
  6. [Section 4.3] The statement that ROC AUC scores in Table 6 do not accurately reflect model performance due to class imbalance is imprecise; ROC AUC is generally invariant to class imbalance, whereas PR AUC is not. Please clarify the intended meaning.
  7. [Throughout] The text uses 'A V' with a space in several places; please use 'AV' consistently.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: EMBER2024's claims are dataset properties and benchmark measurements, not predictions reduced to their inputs.

full rationale

The paper's central claims are descriptive and empirical: it constructs a dataset, describes its labels, and reports benchmark classifier scores. There is no derivation chain in which a fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' prior work, and no ansatz is smuggled in via self-citation. The challenge set is defined by VirusTotal detection status (initially undetected, later detected), and the benign/malicious labels use the same AV-detection criterion; this makes the challenge set a selected subpopulation, not a circular prediction. The statement that classifiers struggle on the challenge set is a measurement against that selection, not an equivalence forced by construction. The family and behavior labels come from the authors' own ClarA Vy tool, which is a disclosed self-citation; however, the paper does not argue for the labels' correctness by citing itself, and the benchmark tasks measure classifier agreement with those labels rather than deriving the labels from the classifiers. Similarly, the near-duplicate exclusion being limited to same-week files is a possible leakage limitation but not a circularity, because it does not make the evaluation score equal to the construction criterion by definition. Overall, the novelty claims and benchmark results are self-contained and externally checkable; the only adjacent concerns are validity risks, not circular reasoning.

Assumptions & free parameters 5 free parameters · 4 assumptions · 2 invented entities

The dataset depends on several design choices: AV detection thresholds, rescan delays, TLSH deduplication, and weekly quotas. Most are disclosed and reasonable, but they introduce noise and selection effects. The two invented entities, the challenge set and feature version 3, are concrete artifacts with independent verification paths through the released code and metadata.

free parameters (5)
  • Malicious AV detection threshold = 5 or more independent AV detections
    Section 2.4: files detected by five or more AV products without known relationships are labeled malicious; this threshold determines all malicious labels and challenge set membership.
  • Benign AV detection threshold = 0 detections after re-scan at 30+ days
    Section 2.4: files with no AV detection after re-scanning are labeled benign; this likely includes never-detected malware.
  • TLSH near-duplicate distance threshold = 30
    Section 2.1: TLSH distance <= 30 treats files as near-duplicates; the threshold is taken from Oliver et al. and affects sample diversity and split integrity.
  • Re-scan delay for label stabilization = 30 days minimum for benign files, 90 days for all files
    Section 2: files are re-queried 90 or more days after first submission; benign suspects are re-scanned at least 30 days later; this schedule defines the challenge set and label stability.
  • Weekly file quotas per format = Win32 30k, Win64 10k, .NET 5k, APK 4k, ELF 500, PDF 1k per week
    Table 2: quotas chosen by availability determine dataset composition and class balance across the 64 collection weeks.
assumptions (4)
  • domain assumption VirusTotal AV detection counts are a valid proxy for file maliciousness
    Section 2.4 uses AV detection counts to define both malicious and benign labels and the challenge set; there is no independent ground truth.
  • domain assumption ClarA Vy family and tag labels are sufficiently accurate for benchmarking
    Section 2.4 uses ClarA Vy to assign family, behavior, packer, exploit, and threat-group labels; any ClarA Vy errors transfer to the benchmark tasks.
  • domain assumption Files first submitted to VirusTotal are representative of malware and benign files encountered in practice
    Section 2 selects files by first submission date to VirusTotal, which is a convenience sample and may not represent all deployment contexts.
  • standard math TLSH distance of 30 identifies near-duplicates with negligible false positive rate
    Section 2.1 relies on Oliver et al. for the threshold; if the false positive rate is higher than reported, duplicate files may leak across splits.
invented entities (2)
  • EMBER2024 challenge set independent evidence
    purpose: A subset of malicious files initially undetected by all AV products, designed to evaluate classifiers against evasive malware.
    Section 2.3 defines the challenge set as a novel evaluation artifact; released feature vectors and construction code allow independent checks of its properties.
  • EMBER feature version 3 independent evidence
    purpose: A unified feature format supporting six file formats, adding DOS header, Rich header, Authenticode, data directory, and parse warning features, plus partial non-PE support.
    Section 3.2 defines the format; the code and vectorized features are released, so the feature values can be independently computed and verified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EMBER2024 -- A Benchmark Dataset for Holistic Evaluation of Malware Classifiers." pith.science (2026). https://pith.science/paper/KQFLU3F3

@misc{pith2026250605074,
  author       = {Pith},
  title        = {Pith review of: EMBER2024 -- A Benchmark Dataset for Holistic Evaluation of Malware Classifiers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KQFLU3F3}},
  note         = {Machine review of arXiv:2506.05074}
}
read the original abstract

A lack of accessible data has historically restricted malware analysis research, and practitioners have relied heavily on datasets provided by industry sources to advance. Existing public datasets are limited by narrow scope - most include files targeting a single platform, have labels supporting just one type of malware classification task, and make no effort to capture the evasive files that make malware detection difficult in practice. We present EMBER2024, a new dataset that enables holistic evaluation of malware classifiers. Created in collaboration with the authors of EMBER2017 and EMBER2018, the EMBER2024 dataset includes hashes, metadata, feature vectors, and labels for more than 3.2 million files from six file formats. Our dataset supports the training and evaluation of machine learning models on seven malware classification tasks, including malware detection, malware family classification, and malware behavior identification. EMBER2024 is the first to include a collection of malicious files that initially went undetected by a set of antivirus products, creating a "challenge" set to assess classifier performance against evasive malware. This work also introduces EMBER feature version 3, with added support for several new feature types. We are releasing the EMBER2024 dataset to promote reproducibility and empower researchers in the pursuit of new malware research topics.

Figures

Figures reproduced from arXiv: 2506.05074 by the authors.

Figure 1
Figure 1. Example VirusTotal report contents. VirusTotal re [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Histogram showing the distribution of family sizes in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Example JSON object displaying a file’s hashes, labels, [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: ROC curve (log scale) of a LightGBM classifier trained on the EMBER2024 training set and evaluated using the EM￾BER2024 test set. The model has a true positive rate of 94.48% when permitting a 1% false positive rate. 4.1 Evaluating a Malware Detection Classifier We tra…
Figure 5
Figure 5. Figure 5: Example of EMBER feature version 3 raw features. [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Most common families and tags in the EMBER2024 dataset. [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Cross-Layer Misalignment Detection in Agent Skills: A Progressive Loading-Aware Contrastive Learning Approach

    cs.AI 2026-07 conditional novelty 6.0 of 10

    PL-HCL detects cross-layer misalignment in Agent Skills by learning consistency among metadata, instructions, and resources, lifting Macro-F1 to 0.87–0.89 on a human-verified challenge set.

  2. Latent Stability Analysis of Malware Representations Under Feature-Space Perturbations

    cs.CR 2026-07 conditional novelty 5.5 of 10

    Latent-stability metrics (LED and PINNFlow shifts) diagnose malware representation movement under feature-space probes, without beating full EMBER or PCA-64 on clean classification.

Reference graph

Works this paper leans on

32 extracted references · 25 canonical work pages · cited by 2 Pith papers

  1. [1]

    [n. d.]. VirusShare.com - Because Sharing is Caring. https://virusshare.com/, Last accessed on 2025-02-17

    work page 2025

  2. [2]

    Hyrum S Anderson and Phil Roth. 2018. Ember: an open dataset for training static pe malware machine learning models. arXiv preprint arXiv:1804.04637 (2018)

    arXiv 2018

  3. [3]

    Apruzzese, H

    G. Apruzzese, H. S. Anderson, S. Dambra, D. Freeman, F. Pierazzi, and K. Roundy

    work page

  4. [4]

    Marcus Botacin, Fabricio Ceschin, Paulo De Geus, and André Grégio. 2020. We need to talk about antiviruses: challenges & pitfalls of av evaluations. Computers & Security 95 (2020), 101859

    work page 2020

  5. [5]

    Ero Carrera. 2004. pefile: Python module to read and work with PE files. https: //github.com/erocarrera/pefile

    work page 2004

  6. [6]

    William W Cohen. 1995. Fast effective rule induction. In Proceedings of the twelfth international conference on machine learning. 115–123

    work page 1995

  7. [7]

    Colin Galen and Robert Steele. 2020. Evaluating performance maintenance and deterioration over time of machine learning-based malware detection models on the ember pe dataset. In 2020 Seventh International Conference on Social Networks Analysis, Management and Security (SNAMS). IEEE, 1–7

    work page 2020

  8. [8]

    Richard Harang and Ethan M Rudd. 2020. SOREL-20M: A large scale benchmark dataset for malicious PE detection. arXiv preprint arXiv:2012.07634 (2020)

    arXiv 2020

Show all 32 references
  1. [9]

    Wenyi Huang and Jack W Stokes. 2016. MtNet: a multi-task neural network for dynamic malware classification. In Detection of Intrusions and Malware, and Vulnerability Assessment: 13th International Conference, DIMVA 2016, San Sebastián, Spain, July 7-8, 2016, Proceedings 13. Sp...

    2016

  2. [10]

    Dash, Zhi Wang, Davide Papini, Ilia Nouretdinov, and Lorenzo Cavallaro

    Roberto Jordaney, Kumar Sharad, Santanu K. Dash, Zhi Wang, Davide Papini, Ilia Nouretdinov, and Lorenzo Cavallaro. 2017. Transcend: Detecting Concept Drift in Malware Classification Models. In 26th USENIX Security Symposium (USENIX Security 17). USENIX Association, Vancouver, ...

    2017

  3. [11]

    Robert J Joyce, Dev Amlani, Charles Nicholas, and Edward Raff. 2023. Motif: A malware reference dataset with ground truth family labels. Computers & Security 124 (2023), 102921

    2023

  4. [12]

    Joyce, Derek Everett, Maya Fuchs, Edward Raff, and James Holt

    Robert J. Joyce, Derek Everett, Maya Fuchs, Edward Raff, and James Holt. 2025. ClarA Vy: A Tool for Scalable and Accurate Malware Family Labeling. InCompan- ion of the 16th ACM/SPEC International Conference on Performance Engineering (WWW Companion ’25)

    2025

  5. [13]

    Robert J Joyce, Edward Raff, Charles Nicholas, and James Holt. 2023. MalDICT: Benchmark Datasets on Malware Behaviors, Platforms, Exploitation, and Packers. arXiv preprint arXiv:2310.11706 (2023)

    2023 arXiv

  6. [14]

    Joseph, and J

    Alex Kantchelian, Michael Carl Tschantz, Sadia Afroz, Brad Miller, Vaishaal Shankar, Rekha Bachwani, Anthony D. Joseph, and J. D. Tygar. 2015. Better Malware Ground Truth: Techniques for Weighting Anti-Virus Vendor Labels. In Proceedings of the 8th ACM Workshop on Artificial I...

    2015

  7. [15]

    Jeremy Z Kolter and Marcus A Maloof. 2004. Learning to detect malicious executables in the wild. In Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 470–478

    2004

  8. [16]

    Nicola Loi, Claudio Borile, and Daniele Ucci. 2021. Towards an automated pipeline for detecting and classifying malware through machine learning. arXiv preprint arXiv:2106.05625 (2021)

    2021 arXiv

  9. [17]

    Microsoft Security Team. 2021. Attack AI Systems in Machine Learning Evasion Competition. https://www.microsoft.com/en-us/security/blog/2021/07/29/attack- ai-systems-in-machine-learning-evasion-competition/

    2021

  10. [18]

    Aziz Mohaisen, Omar Alrawi, and Manar Mohaisen. 2015. AMAL: High-fidelity, behavior-based automated malware analysis and classification. Computers & Security 52 (2015), 251 – 266. doi:10.1016/j.cose.2015.04.001

    2015 doi

  11. [19]

    Jonathan Oliver, Chun Cheng, and Yanggui Chen. 2013. TLSH–a locality sensitive hash. In 2013 Fourth Cybercrime and Trustworthy Computing Workshop. IEEE, 7–13

    2013

  12. [20]

    Tirth Patel, Fred Lu, Edward Raff, Charles Nicholas, Cynthia Matuszek, and James Holt. 2023. Small Effect Sizes in Malware Detection? Make Harder Train/Test Splits! (2023). EMBER2024 - A Benchmark Dataset for Holistic Evaluation of Malware Classifiers KDD’25, August 3–7, 2025,...

    2023

  13. [21]

    Edward Raff and Charles Nicholas. 2020. A survey of machine learning methods and challenges for windows malware classification. arXiv preprint arXiv:2006.09271 (2020)

    2020 arXiv

  14. [22]

    Phil Roth. 2019. EMBER Improvements. (2019). https://docs.google.com/ presentation/d/1A13tsUkgWeujTy9SD-vDFfQp9fnIqbSE_tCihNPlArQ Confer- ence on Applied Machine Learning in Information Security

    2019

  15. [23]

    Joshua Saxe and Konstantin Berlin. 2015. Deep neural network based malware detection using two dimensional binary program features. In Malicious and Un- wanted Software (MALWARE), 2015 10th International Conference on . IEEE, 11–20

    2015

  16. [24]

    Matthew G Schultz, Eleazar Eskin, F Zadok, and Salvatore J Stolfo. 2001. Data mining methods for detection of new malicious executables. In Security and Privacy, 2001. S&P 2001. Proceedings. 2001 IEEE Symposium on. IEEE, 38–49

    2001

  17. [25]

    Marcos Sebastián, Richard Rivera, Platon Kotzias, and Juan Caballero. 2016. Av- class: A tool for massive malware labeling. InResearch in Attacks, Intrusions, and Defenses: 19th International Symposium, RAID 2016, Paris, France, September 19-21, 2016, Proceedings 19. Springer, 230–253

    2016

  18. [26]

    John Seymour. [n. d.]. label-virusshare. https://github.com/seymour1/label- virusshare, Last accessed on 2025-02-14

    2025

  19. [27]

    Wei Song, Xuezixiang Li, Sadia Afroz, Deepali Garg, Dmitry Kuznetsov, and Heng Yin. 2021. MAB-Malware: A Reinforcement Learning Framework for Attacking Static Malware Classifiers. arXiv:2003.03100 [cs.CR] https://arxiv.org/ abs/2003.03100

    2021 arXiv

  20. [28]

    Romain Thomas. 2017. LIEF - Library to Instrument Executable Formats. https://lief.quarkslab.com/

    2017

  21. [29]

    Xabier Ugarte-Pedrero, Mariano Graziano, and Davide Balzarotti. 2019. A close look at a daily dataset of malware samples. ACM Transactions on Privacy and Security (TOPS) 22, 1 (2019), 1–30

    2019

  22. [30]

    VirusTotal. [n. d.]. Analyse suspicious files, domains, IPs and URLs to detect mal- ware and other breaches, automatically share them with the security community. https://www.virustotal.com/en/, Last accessed on 2025-02-18

    2025

  23. [31]

    histogram

    Shao-Huai Zhang, Cheng-Chung Kuo, and Chu-Sing Yang. 2019. Static PE Malware Type Classification Using Machine Learning Techniques. In 2019 In- ternational Conference on Intelligent Computing and its Emerging Applications (ICEA). 81–86. doi:10.1109/ICEA.2019.8858297 KDD’25, Au...

    2019

  24. [2023]

    Real attackers don’t compute gradients

    “Real attackers don’t compute gradients”: Bridging the gap between ad- versarial ML research and practice. In 2023 IEEE Conference on Secure and Trustworthy Machine Learning (SaTML). IEEE, 339–364

    2023

Pith tools

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