Pith. sign in

REVIEW 2 major objections 3 minor 49 references

A Comparison of Malware Image Transformations Using Grad-CAM and Hybrid Learning Models

T0 review · 2 major / 3 minor · reviewed 2026-08-16 · deepseek-v4-flash

Pith's one-line read The paper claims that combining CNN embeddings from original malware images and their Grad-CAM overlays achieves 0.777 test accuracy across 17 malware families, beating the previous 0.751 benchmark, and that accuracy and explanation…

desk verdict The headline 0.777 accuracy is invalid: the Grad-CAM overlays are generated with the true family label as target, so the features leak the answer — but the faithfulness/stability comparison across transformations is a genuinely useful contribution that deserves a serious referee. read the letter →

arxiv 2608.12077 v1 pith:QXUWUAWW submitted 2026-08-12 cs.CR

classification cs.CR
keywords image-basedmalwareclassificationGrad-CAMexplainableAIMobileNetV2RandomForestfamilyfaithfulnessmetricsbinary-to-imagetransformations
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish two things: that Grad-CAM heatmaps of malware images carry enough family-discriminative structure to be used as classification features, and that the choice of binary-to-image transformation affects accuracy and explanation quality in different ways. On the first point, it reports that a Random Forest trained on a 4096-dimensional vector of MobileNetV2 embeddings, built from original images and Grad-CAM overlays across eight transformations, reaches 0.777 test accuracy on 17 malware families, above the previous best of 0.751 on the same dataset. On the second point, it reports that the most faithful explanations come from Bigram Polar, a mid-tier accuracy transformation, while the most accurate transformation, Entropy Hilbert, is also second-most faithful and therefore the best single compromise. A sympathetic reader would care because the result suggests explanations are not just a diagnostic add-on but a usable signal for classification, and because the accuracy/faithfulness split warns against choosing an image representation by accuracy alone.

What carries the argument

The load-bearing object is the Grad-CAM heatmap, $L^c_{\mathrm{Grad\text{-}CAM}} = \mathrm{ReLU}\bigl(\sum_k \alpha_k^c A^k\bigr)$, where $A^k$ is the last convolutional activation map of MobileNetV2 and $\alpha_k^c$ is the spatially averaged gradient of the class logit with respect to $A^k$. The heatmap is generated for every sample and transformation, and the true family label is used as the target class. Three uses carry the argument: overlaying the heatmap on the image to create a new input representation, extracting fifty handcrafted statistical, shape, and texture features from the raw heatmap for classical classifiers, and using the CNN's 256-dimensional penultimate-layer embedding from each original and overlay model. The final model concatenates all sixteen 256-dimensional embeddings into a 4096-dimensional vector and classifies it with Random Forest. Faithfulness is quantified as insertion AUC minus deletion AUC from pixel-perturbation experiments, and stability is quantified by Spearman correlation, SSIM, and top-20% overlap under Gaussian noise.

What would settle it

Regenerate the overlay dataset with the model's predicted class as the Grad-CAM target, retrain the 4096-dimensional Random Forest, and compare test accuracy to 0.777; a substantial drop would show the headline result depends on label leakage into feature generation.

Watch

Extended reading notes

Core claim

The central claim is that a Random Forest trained on concatenated penultimate-layer embeddings from 16 MobileNetV2 models—eight trained on original transformation images and eight on their Grad-CAM overlays—attains 0.777 test accuracy on the 17-family dataset, exceeding the 0.751 benchmark from prior work on the same data. The paper also claims that accuracy and explanation faithfulness do not coincide: Bigram Polar ranks first on faithfulness but seventh on CNN accuracy, whereas Grayscale ranks well on accuracy but seventh on faithfulness, and Entropy Hilbert is the only transformation that is strong on both. In addition, Grad-CAM overlays are claimed to be more than visualizations: for five of the eight transformations, CNNs trained on overlay images match or beat those trained on the original images, with the largest gains for Grayscale and Byteclass Hilbert.

Load-bearing premise

The pipeline assumes the true family label is available when generating each Grad-CAM overlay, because the explanation is computed with the ground-truth class as target; if a deployed system must produce overlays without knowing the label, the reported 0.777 accuracy describes a different, untested procedure.

Editorial extensions

If this is right

  • Explanations themselves carry family-discriminative signal: a 4096-dimensional Random Forest built from CNN embeddings of original and Grad-CAM overlay images reaches 0.777 test accuracy on 17 families, above the previous 0.751 benchmark on the same dataset.
  • For five of eight transformations, CNNs trained on Grad-CAM overlay images match or outperform CNNs trained on the original images, so overlays can serve as an input representation rather than only a diagnostic tool.
  • Accuracy alone is a poor guide to explainability: the most faithful transformation, Bigram Polar, is mid-tier in accuracy, while the most accurate transformation, Entropy Hilbert, is also second-most faithful, making it the best compromise.
  • Faithfulness and stability are inversely ranked across transformations, so explanation quality is not a single axis; an analyst must choose between explanations that are correct and explanations that are robust to small input changes.

Reading between the lines

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

  • The paper leaves implicit that its central comparison would change under a predicted-label protocol: if Grad-CAM targets the model's own predicted class instead of the true label, the overlay and 4096-dimensional accuracies could drop, and that drop would measure how much label information leaks into feature generation.
  • A natural extension is to add HOG features to the 4096-dimensional vector; the paper shows HOG consistently helps in smaller hybrids, so the same gain may appear at full fusion, though the paper does not test it.
  • The faithfulness-stability inversion suggests a practical selection rule for security analysts: choose Entropy Hilbert for a single representation that serves both accuracy and human review, and treat Bigram Polar explanations as precise but fragile.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 3 minor

Summary. The manuscript compares eight binary-to-image transformations for malware family classification on a 17-class, 17,000-sample dataset. It trains MobileNetV2 CNNs on original and Grad-CAM overlay images, extracts handcrafted features from Grad-CAM heatmaps, builds hybrid CNN-HOG-XGBoost pipelines, and concatenates CNN embeddings into 256-, 512-, and 4096-dimensional feature vectors. The headline result is a 0.777 test accuracy for a Random Forest on the 4096-dimensional embeddings, which the authors claim exceeds a prior benchmark of 0.751. The paper also evaluates faithfulness and stability of Grad-CAM explanations, compares Grad-CAM with HiResCAM, and concludes that accuracy and explanation faithfulness do not coincide across the eight transformations.

Significance. If valid, the paper would offer a useful systematic comparison of image transformations for malware classification and an empirical study of Grad-CAM behavior in this domain. The faithfulness/stability analysis and the HiResCAM comparison are of interest, and the authors disclose practical details such as fixed seeds and three excluded degenerate heatmaps. However, the central classification benchmark is invalid because the Grad-CAM overlays are generated using the true family label as the target class, so the input features encode the label the classifier is supposed to predict. This undermines the headline accuracy claims and the accuracy component of the accuracy-versus-faithfulness finding.

major comments (2)
  1. [3.2, 4.3-4.6, Tables 3, 4, 8, 9] Section 3.2 states: 'The true family label is used as the target class c, so that the explanation corresponds to the ground-truth class.' The resulting overlay images are then used as inputs to the overlay CNNs (Table 3), the CNN-HOG-XGBoost hybrid (Table 4), and the 256/512/4096-dimensional embeddings (Tables 8 and 9). Because the test overlays are also generated using each test sample's true family label, the input features carry the label information the classifier is supposed to predict. The reported test accuracies, including the 0.777 in the Abstract and Table 9, therefore do not describe a deployable classifier; they measure label-conditioned representations. This also invalidates the accuracy component of the 'accuracy and explanation faithfulness do not coincide' conclusion in Sections 4.5 and 5.
  2. [4.6, Table 9] The 4096-dimensional experiment does not specify the train/test split used for the final Random Forest. The text reports only that samples present across all 16 models are included (16,997 samples) and that a Random Forest with 500 trees is trained. Since Section 3.1 constructs per-transformation stratified splits separately, it is unclear how these splits are reconciled when samples are pooled across all transformations. Without a clean single split, the 0.777 accuracy could be further inflated even beyond the leakage problem identified in the previous comment.
minor comments (3)
  1. [Abstract] The Abstract reports a previous benchmark of 0.750, while Sections 4.6 and Table 11 report 0.751 from Agrawal et al. [2]; the abstract should be corrected to 0.751 or the discrepancy should be explained.
  2. [Section 6] In the first paragraph of Section 6, 'There are several promising directions for extending the work considered in this chapter, We showed that combining HOG descriptors...' should read 'chapter. We showed' with a period instead of a comma.
  3. [Tables 3, 4, and 7] Overlay-image CNN accuracies appear with different values for the same transformations across Tables 3, 4, and 7 (e.g., Grayscale 0.566, 0.565, and 0.602). The paper should clarify which of these runs correspond to the same or different training conditions, since the current presentation makes the discrepancies difficult to interpret.

Circularity Check

2 steps flagged · score 7.0 of 10

Grad-CAM overlays are constructed from the true family label (Section 3.2), then used as classifier inputs; the reported 0.777 Random Forest accuracy is a label-leaked benchmark rather than a valid prediction.

  1. self definitional [Section 3.2 (Grad-CAM definition); applied in Sections 4.3-4.6 and Tables 3, 4, 7-9]
    "The true family label is used as the target class c, so that the explanation corresponds to the ground-truth class. ... In addition to using Grad-CAM as a visualization tool, we also use the Grad-CAM heatmaps as input for further analysis. Grad-CAM maps are generated for each of our malware image representations. Features are then extracted from these heatmaps, and classical machine learning models are trained on the resulting feature vectors. For comparison, CNNs are also trained directly on the Grad-CAM images."

    Grad-CAM's map is defined as L^c = ReLU(sum_k alpha^c_k A^k), where the class logit y^c is chosen with c equal to the true family label. The resulting overlay is then treated as ordinary input data for CNNs and for feature extraction. Because the overlay image is a function of the ground-truth label, both training and test overlays carry label information by construction. At inference time, constructing the test overlay requires knowing the test sample's true family, so the reported classification accuracies do not describe a classifier usable on label-free malware inputs.

  2. self definitional [Abstract; Section 4.6 (Table 9)]
    "Specifically, we show that a Random Forest model trained on features extracted from Grad-CAM images via a MobileNetV2 Convolutional Neural Network (CNN) model achieves a test accuracy of 0.777 across 17 malware families, exceeding a previous benchmark of 0.750 for this same dataset."

    The headline 0.777 result is the Random Forest test accuracy on 4096-dimensional vectors formed by concatenating 256-dimensional embeddings from all 16 models, including models trained on Grad-CAM overlays (Section 4.6). Since each overlay was generated using the true family label as Grad-CAM's target class (Section 3.2), these embeddings encode the ground-truth label at both training and test time. The 'test accuracy' therefore measures how well the model recognizes labels that were already used to construct its inputs, not how well it classifies unseen binaries from label-free images.

full rationale

The paper's central classification benchmark is circular at the experimental-protocol level rather than at the equation level. Section 3.2 defines Grad-CAM with the true family label as target class c, which is a standard choice for explanation faithfulness evaluation. However, the paper then repurposes these label-conditioned overlays as ordinary input images and as the source of CNN embeddings for downstream classifiers (Sections 4.3, 4.4, and 4.6). Because the overlays are generated from the ground-truth label, every train and test overlay already contains information about the target family; the 0.777 Random Forest accuracy (Abstract, Table 9) is therefore a label-leaked benchmark rather than a valid prediction from malware binaries alone. The same leakage affects the overlay CNN accuracies in Tables 3, 4, 7, and 8 and the 256- and 512-dimensional embedding results. The faithfulness and stability analyses (Tables 5 and 6) are separate and not circular, since using the true class as the explanation target is the accepted protocol for those metrics, and the faithfulness-versus-accuracy comparison is based on original-image CNN accuracies. No load-bearing self-citation chain is present; the comparison with the prior 0.750 benchmark in [2] is an external benchmark, not a derivational input. Because the central claim reduces to label-conditioned inputs, but the XAI analyses retain independent content, a score of 7 is appropriate.

Assumptions & free parameters 13 free parameters · 7 assumptions · 0 invented entities

The central result depends on a long list of hand-chosen hyperparameters, the correctness of the public malware labels, the fidelity of the eight transformations to prior work, and the label-bearing Grad-CAM overlay protocol. No new physical or conceptual entities are introduced.

free parameters (13)
  • dropout_rate = 0.3
    Dropout after global average pooling in classification head is hand-chosen; no tuning curve is reported.
  • dense_layer_size = 256
    Penultimate dense layer dimension used as embedding; hand-chosen.
  • learning_rates = 0.001 then 0.0001
    Two-stage training schedule is a modeling choice, not derived from data.
  • fine_tune_fraction = upper 30 percent of backbone layers
    Amount of backbone unfrozen is chosen by hand.
  • random_forest_trees = 500
    Classifier hyperparameter for the final 4096-dimensional experiment.
  • hog_parameters = 9 orientation bins, 8x8 cells, 2x2 blocks
    HOG descriptor settings are chosen; no sensitivity analysis.
  • cam_mask_threshold = 0.5
    Threshold for deriving binary mask and blob features from CAM.
  • glcm_thresholds = 0.3 and 0.7
    Two CAM thresholds used to define low and high attention regions for GLCM texture features.
  • faithfulness_perturbation_fractions = 0.1 through 1.0
    Discretization of deletion/insertion curves is a methodological choice.
  • noise_sigma = 5 on a 0-255 scale
    Stability perturbation magnitude chosen by hand.
  • overlay_alpha = 0.45
    Blending factor for heatmap overlay images.
  • svm_C = 10
    RBF SVM regularization hyperparameter for 256/512-dimensional experiments.
  • xgboost_learning_rate = 0.05
    Hyperparameter for XGBoost in feature fusion experiments.
assumptions (7)
  • domain assumption The 17 family labels in the RawMalTF-derived dataset are correct and the 1,000 samples per family are representative.
    All accuracy numbers depend on ground-truth labels as given by [5] and [2].
  • domain assumption Converting the first 50,176 bytes of each binary into an image preserves enough family-discriminative structure for classification.
    The entire study relies on this premise from [2]; used in Section 2.1.
  • ad hoc to paper The eight image transformations are implemented faithfully as in Agrawal et al. [2].
    No code is shipped, so the comparison inherits the prior implementation without verification.
  • domain assumption MobileNetV2 with ImageNet weights is an appropriate feature extractor for malware imagery.
    The backbone is fixed a priori; Section 3.1.
  • domain assumption Deletion and insertion AUC faithfully measure explanation faithfulness.
    Adopted from [19,20,33] without validating the metrics on malware images.
  • ad hoc to paper True family labels are available when generating Grad-CAM overlays used as classifier inputs.
    Section 3.2 states 'The true family label is used as the target class c.' This is the label-leakage assumption.
  • ad hoc to paper Per-transformation stratified splits remain valid when samples are pooled into the 4096-dimensional feature set.
    Section 4.6 pools 16,997 samples across models trained with separately drawn splits but does not define the final train/test split.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Comparison of Malware Image Transformations Using Grad-CAM and Hybrid Learning Models." pith.science (2026). https://pith.science/paper/QXUWUAWW

@misc{pith2026260812077,
  author       = {Pith},
  title        = {Pith review of: A Comparison of Malware Image Transformations Using Grad-CAM and Hybrid Learning Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QXUWUAWW}},
  note         = {Machine review of arXiv:2608.12077}
}
read the original abstract

Recent studies have shown that binary-to-image representations can enable effective machine learning-based results for malware detection and classification. However, performance can vary significantly, depending on the technique used to convert binaries to images. Furthermore, the explainability and interpretability of image-based models is largely unexplored within the malware domain. In this research, we employ Gradient-weighted Class Activation Maps (Grad-CAM) as an eXplainable AI (XAI) tool, which we use to analyze eight distinct image types derived from malware samples. We provide quantitative faithfulness and stability metrics for Grad-CAM heatmaps and we compare these heatmaps to High-Resolution Class Activation Mappings (HiResCAM). We also show that Grad-CAM heatmaps can provide useful information for malware classification. Specifically, we show that a Random Forest model trained on features extracted from Grad-CAM images via a MobileNetV2 Convolutional Neural Network (CNN) model achieves a test accuracy of 0.777 across 17 malware families, exceeding a previous benchmark of 0.750 for this same dataset. A key finding of this research is that for the malware image transformations considered, accuracy and explanation faithfulness do not coincide, e.g., image transformation techniques that produce the most faithful explanations yield only mid-tier accuracy.

Figures

Figures reproduced from arXiv: 2608.12077 by the authors.

Figure 1
Figure 1. Two examples of each image transformation type [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Malware images and corresponding Grad-CAM overlays [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Comparison of Grad-CAM and HiResCAM heatmaps [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Confusion matrices all transformations in both settings. Furthermore, the 512-dimensional combined features improve over the 256-dimensional overlay-only features in every transforma￾tion, confirming that the original image model captures complementary information to t…
Figure 5
Figure 5. Figure 5: Confusion matrix for 4096-dimensional features (accuracy = 0 [PITH_FULL_IMAGE:figures/full_fig_p023_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 43 canonical work pages

  1. [1]

    Agensla: Trojan-PSW.MSIL.Agensla.https://threats.kaspersky.com/en/ threat/Trojan-PSW.MSIL.Agensla/, 2025

  2. [2]

    A comparison of selected image transformation techniques for malware classification

    Rishit Agrawal, Kunal Bhatnagar, Andrew Do, Ronnit Rana, Martin Jureˇ cek, and Mark Stamp. A comparison of selected image transformation techniques for malware classification. In Roberto Di Pietro, Karen Renaud, and Paolo Mori, editors,Proceedings of the 12th International Conference on Information Systems Security and Privacy, volume 2 ofICISSP, pages 33...

  3. [3]

    Androm: Backdoor.Win32.Androm.https://threats.kaspersky.com/en/ threat/Backdoor.Win32.Androm/, 2025

  4. [4]

    Springer, 2006

    John Aycock.Computer Viruses and Malware. Springer, 2006

  5. [5]

    RawMal-TF: Raw malware dataset labeled by type and family.https://arxiv.org/abs/2506.23909, 2025

    David B´ alik, Martin Jureˇ cek, and Mark Stamp. RawMal-TF: Raw malware dataset labeled by type and family.https://arxiv.org/abs/2506.23909, 2025

  6. [6]

    Low, and Mark Stamp

    Donabelle Baysa, Richard M. Low, and Mark Stamp. Structural entropy and metamorphic malware.Journal of Computer Virology and Hacking Techniques, 9(4):179–192, 2013

  7. [7]

    Trans- fer learning for image-based malware classification

    Niket Bhodia, Pratikkumar Prajapati, Fabio Di Troia, and Mark Stamp. Trans- fer learning for image-based malware classification. In Paolo Mori, Steven Fur- nell, and Olivier Camp, editors,Proceedings of the 5th International Conference on Information Systems Security and Privacy, ICISSP, pages 719–726, 2019

  8. [8]

    Random forests.Machine learning, 45(1):5–32, 2001

    Leo Breiman. Random forests.Machine learning, 45(1):5–32, 2001

Show all 49 references
  1. [9]

    Generalizing adversarial explanations with Grad- CAM

    Tanmay Chakraborty et al. Generalizing adversarial explanations with Grad- CAM. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, CVPRW, pages 187–193, 2022. 27

  2. [10]

    XGBoost: A scalable tree boosting system

    Tianqi Chen and Carlos Guestrin. XGBoost: A scalable tree boosting system. http://arxiv.org/abs/1603.02754, 2016

  3. [11]

    Dynamic malware detection based on enhanced semantic API sequence features.Expert Systems with Applications, 315:131781, 2026

    Zhiguo Chen, Lei Zhou, Qingcheng Liu, Weizhi Meng, and Jian Weng. Dynamic malware detection based on enhanced semantic API sequence features.Expert Systems with Applications, 315:131781, 2026

  4. [12]

    Convagent: Trojan.Win32.CONVAGENT.0NA103AQ24.https: //www.trendmicro.com/vinfo/us/threat-encyclopedia/malware/ Trojan.Win32.CONVAGENT.0NA103AQ24, 2024

  5. [13]

    Crypt: HackTool.Win32.Crypt.au.https://threats.kaspersky.com/en/ threat/HackTool.Win32.Crypt.au/, 2025

  6. [14]

    Crysan: Backdoor.Win32.Crysan.gen.https://threats.kaspersky.com/en/ threat/VHO:Backdoor.Win32.Crysan.gen/, 2025

  7. [15]

    Histograms of oriented gradients for human detection

    Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection. In2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, volume 1 ofCVPR’05, pages 886–893, 2005

  8. [16]

    Austin, and Mark Stamp

    Anusha Damodaran, Fabio Di Troia, Corrado Aaron Visaggio, Thomas H. Austin, and Mark Stamp. A comparison of static, dynamic, and hybrid analysis for malware detection.Journal of Computer Virology and Hacking Techniques, 13(1):1–12, 2017

  9. [17]

    DCRat: Backdoor.MSIL.DCRat.aae.https://threats.kaspersky.com/en/ threat/Backdoor.MSIL.DCRat.aae/, 2025

  10. [18]

    Use HiResCAM instead of Grad- CAM for faithful explanations of convolutional neural networks.https:// arxiv.org/abs/2011.08891, 2020

    Rachel Lea Draelos and Lawrence Carin. Use HiResCAM instead of Grad- CAM for faithful explanations of convolutional neural networks.https:// arxiv.org/abs/2011.08891, 2020

  11. [19]

    Computing and evaluating saliency maps for image classification: a tutorial.Journal of Electronic Imaging, 32(2):020801, 2023

    Tristan Gomez and Harold Mouch` ere. Computing and evaluating saliency maps for image classification: a tutorial.Journal of Electronic Imaging, 32(2):020801, 2023

  12. [20]

    Naofumi Hama, Masayoshi Mase, and Art B. Owen. Deletion and insertion tests in regression models.Journal of Machine Learning Research, 24:1–38, 2023

  13. [21]

    Haralick, K

    Robert M. Haralick, K. Shanmugam, and Its’Hak Dinstein. Textural features for image classification.IEEE Transactions on Systems, Man, and Cybernetics, SMC-3(6):610–621, 1973

  14. [22]

    Injuke: Trojan.Win32.Injuke.gen.https://threats.kaspersky.com/en/ threat/HEUR:Trojan.Win32.Injuke.gen/, 2025

  15. [23]

    Malware detection with structural entropy features using multilayer perceptron neural network

    Yeong Tyng Ling, Piau Phang, Kang Leng Chiew, and Xiaowei Zhang. Malware detection with structural entropy features using multilayer perceptron neural network. In2022 International Conference on Digital Transformation and Intelligence, ICDI, pages 01–07, 2022

  16. [24]

    MalwareBazaar.https://bazaar.abuse.ch/, 2026

  17. [25]

    GIWRF-SMOTE: Gini impurity- based weighted random forest with SMOTE for effective malware attack and anomaly detection in IoT-Edge.Smart Science, 11(2):276–292, 2023

    Jadshan Manokaran and Gurusami Vairavel. GIWRF-SMOTE: Gini impurity- based weighted random forest with SMOTE for effective malware attack and anomaly detection in IoT-Edge.Smart Science, 11(2):276–292, 2023. 28

  18. [26]

    Kimmell, Mahmoud Ab- delsalam, Sajad Khorsandroo, and Maanak Gupta

    Harikha Manthena, Shaghayegh Shajarian, Jeffrey C. Kimmell, Mahmoud Ab- delsalam, Sajad Khorsandroo, and Maanak Gupta. Explainable artificial intel- ligence (XAI) for malware analysis: A survey of techniques, applications, and open challenges.IEEE Access, 13:61611–61640, 2025

  19. [27]

    Evaluation of printable character-based malicious PE file- detection method.Internet of Things, 19:100521, 2022

    Mamoru Mimura. Evaluation of printable character-based malicious PE file- detection method.Internet of Things, 19:100521, 2022

  20. [28]

    Mokes: Backdoor.Win32.Mokes.https://threats.kaspersky.com/en/ threat/Backdoor.Win32.Mokes/, 2025

  21. [29]

    Malware images: Visualization and automatic classification

    Lakshmanan Nataraj, Shankarapani Karthikeyan, George Jacob, and BS Man- junath. Malware images: Visualization and automatic classification. InPro- ceedings of the 8th International Symposium on Visualization for Cyber Secu- rity, pages 1–7, 2011

  22. [30]

    Advancing mal- ware imagery classification with explainable deep learning: A state-of-the-art approach using SHAP, LIME and Grad-CAM.PLOS ONE, 20(5):e0318542, 2025

    Sadia Nazim, Muhammad Mansoor Alam, Syed Safdar Rizvi, Jawahir Che Mustapha, Syed Shujaa Hussain, and Mazliham Mohd Suud. Advancing mal- ware imagery classification with explainable deep learning: A state-of-the-art approach using SHAP, LIME and Grad-CAM.PLOS ONE, 20(5):e0318542, 2025

  23. [31]

    Noon: Trojan-Spy.Win32.Noon.https://threats.kaspersky.com/en/ threat/Trojan-Spy.Win32.Noon/, 2025

  24. [32]

    Timo Ojala, Matti Pietik¨ ainen, and Topi M¨ aenp¨ a¨ a. Multiresolution gray-scale and rotation invariant texture classification with local binary patterns.IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(7):971–987, 2002

  25. [33]

    RISE: Randomized input sampling for explanation of black-box models

    Vitali Petsiuk, Abir Das, and Kate Saenko. RISE: Randomized input sampling for explanation of black-box models. InBritish Machine Vision Conference, BMVC, page 151, 2018

  26. [34]

    Remcos: Backdoor.Win32.Remcos.aaaa.https://threats.kaspersky.com/ en/threat/Backdoor.Win32.Remcos.aaaa/, 2025

  27. [35]

    Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra

    Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-CAM: Visual explanations from deep networks via gradient-based localization. In2017 IEEE International Conference on Computer Vision, ICCV, pages 618–626, 2017

  28. [36]

    kaspersky.com/en/threat/HEUR:Trojan-Downloader.Win32.Seraph.gen/, 2025

    Seraph: Trojan-Downloader.Win32.Seraph.gen.https://threats. kaspersky.com/en/threat/HEUR:Trojan-Downloader.Win32.Seraph.gen/, 2025

  29. [37]

    An empirical analysis of invariance Hu’s moment feature over a digi- tal image

    Brajesh Kumar Singh, Amrita Rai, Krishanu Kundu, Karabi Kalita, and Reshu Agrawal. An empirical analysis of invariance Hu’s moment feature over a digi- tal image. In2024 11th International Conference on Reliability, Infocom Tech- nologies and Optimization (Trends and Future Di...

  30. [38]

    kaspersky.com/en/threat/HEUR:Trojan-Spy.Win32.SnakeLogger.gen/, 2025

    Snakelogger: Trojan-Spy.Win32.SnakeLogger.gen.https://threats. kaspersky.com/en/threat/HEUR:Trojan-Spy.Win32.SnakeLogger.gen/, 2025. 29

  31. [39]

    A selective survey of deep learning techniques and their ap- plication to malware analysis

    Mark Stamp. A selective survey of deep learning techniques and their ap- plication to malware analysis. In Mark Stamp, Mamoun Alazab, and Andrii Shalaginov, editors,Malware Analysis Using Artificial Intelligence and Deep Learning, pages 3–51. Springer, 2021

  32. [40]

    com/en/threat/HEUR:Trojan-PSW.Win32.Stealerc.gen/, 2025

    Stealerc: Trojan-PSW.Win32.Stealerc.gen.https://threats.kaspersky. com/en/threat/HEUR:Trojan-PSW.Win32.Stealerc.gen/, 2025

  33. [41]

    Strab: Trojan.Win32.Strab.gen.https://threats.kaspersky.com/en/ threat/HEUR:Trojan.Win32.Strab.gen/, 2025

  34. [42]

    BHMDC: A byte and hex n-gram based malware detection and classification method

    Yonghe Tang, Xuyan Qi, Jing Jing, Chunling Liu, and Weiyu Dong. BHMDC: A byte and hex n-gram based malware detection and classification method. Computers & Security, 128:103118, 2023

  35. [43]

    kaspersky.com/en/threat/HEUR:Trojan-Downloader.Win32.Taskun.gen/, 2025

    Taskun: Trojan-Downloader.Win32.Taskun.gen.https://threats. kaspersky.com/en/threat/HEUR:Trojan-Downloader.Win32.Taskun.gen/, 2025

  36. [44]

    Simple binary data visualization.https://martin.varela

    Mart´ ın Varela. Simple binary data visualization.https://martin.varela. fi/2017/09/09/simple-binary-data-visualization/, 2017

  37. [45]

    VirusShare.https://www.virusshare.com/, 2026

  38. [46]

    VX Underground.https://vx-underground.org/, 2026

  39. [47]

    Hunting for metamorphic engines.Journal in Computer Virology, 2(3):211–229, 2006

    Wing Wong and Mark Stamp. Hunting for metamorphic engines.Journal in Computer Virology, 2(3):211–229, 2006

  40. [48]

    Deep learning versus Gist descriptors for image-based malware classifi- cation

    Sravani Yajamanam, Vikash Raja Samuel Selvin, Fabio Di Troia, and Mark Stamp. Deep learning versus Gist descriptors for image-based malware classifi- cation. In Paolo Mori, Steven Furnell, and Olivier Camp, editors,Proceedings of the 4th International Conference on Information...

  41. [49]

    Appendix In this Appendix, we give examples of malware images and their corresponding Grad-CAM overlay images

    Zenpak: Trojan.Win32.Zenpak.gen.https://threats.kaspersky.com/en/ threat/HEUR:Trojan.Win32.Zenpak.gen/, 2025. Appendix In this Appendix, we give examples of malware images and their corresponding Grad-CAM overlay images. Of the eight image conversion types considered in this c...

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.