Pith. sign in

REVIEW 4 major objections 7 minor 32 references

Splitting a ransomware detector into shards lets you delete data in seconds

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 →

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

T0 review · glm-5.2

2026-07-10 00:03 UTC pith:X5SKH6CV

load-bearing objection Genuine integration of SISA unlearning with DDQN-based ransomware detection, but the favorable privacy result may be an artifact of undertrained retrained shards. the 4 major comments →

arxiv 2607.06860 v1 pith:X5SKH6CV submitted 2026-07-07 cs.CR

Auditable Machine Unlearning for Privacy-Compliant Ransomware Detection Using Multi-Shard SISA and Deep Reinforcement Learning

Jannatul Ferdous , Rafiqul Islam , Md Zahidul Islam This is my paper
classification cs.CR
keywords detectionransomwareremovalretrainingunlearningauditablebehavioraldeep
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.

Privacy laws like GDPR can require organizations to delete specific user data from trained machine learning models. For ransomware detectors trained on behavioral traces, the standard approach—retraining the entire model from scratch after each deletion request—takes minutes and disrupts the security pipeline. This paper proposes splitting the training data into independent partitions called shards, training a separate Double Deep Q-Network detector on each shard, and combining their predictions by majority vote. When a deletion request arrives, only the shard containing the targeted sample is retrained, reducing the operation from 80–330 seconds down to 5–30 seconds. The paper evaluates this system on a balanced dataset of 2,000 Windows 11 behavioral samples across 270 deletion-stress configurations, varying shard counts, deletion fractions, and repeated rounds. The central claim is that with a moderate number of shards (5 to 10), the framework preserves detection accuracy above an F1 of 0.96, keeps the unlearned model's behavior close to that of a from-scratch retrained reference model, and maintains membership-inference attack success near random guessing (0.5), all while cutting retraining cost by an order of magnitude. The paper also shows that pushing shard granularity too high (20 shards) under aggressive repeated deletions causes the ensemble to collapse, as individual shards run out of data.

Core claim

The paper identifies a practical operating regime—5 to 10 shards—where shard-isolated reinforcement learning detectors can be selectively retrained after data deletion requests in 5–30 seconds while preserving F1 above 0.96, maintaining oracle disagreement below 0.04, and keeping membership inference scores near 0.5. The discovery is that this regime exists at all: that splitting a DDQN-based ransomware detector into independent shards does not destroy detection accuracy, and that localized retraining of only the affected shard approximates full retraining closely enough for compliance purposes under moderate deletion pressure. The paper also discovers and characterizes a boundary condition:

What carries the argument

The framework rests on four components. First, a Double Deep Q-Network (DDQN) formulates ransomware detection as a Markov decision process where each behavioral feature vector is a state, the two actions are benign/ransomware, and the reward function is asymmetric—penalizing missed ransomware detections (false negatives) at −2 versus false positives at −0.5. Second, the SISA (Sharded, Isolated, Sliced, Aggregated) framework partitions the training data into M non-overlapping, class-balanced shards, trains an independent DDQN on each, and aggregates predictions via majority voting. Third, an oracle verification protocol compares the selectively unlearned model against a reference model re-re-

Load-bearing premise

The framework assumes that training a separate DDQN on each shard of roughly 200–400 samples produces a Q-value policy that, when aggregated by majority vote, approximates the performance of a single DDQN trained on the full dataset. The paper reports that SISA training already drops F1 from 0.9925 (full data) to 0.953–0.979 (sharded, before any unlearning), but treats this pre-unlearning degradation as a baseline rather than analyzing whether shard isolation fundamentally de

What would settle it

If, on a larger or more heterogeneous dataset, the gap between the sharded-ensemble model and the full-data model widens substantially before any deletion occurs—say, F1 drops by more than 5 percentage points purely from sharding—then the entire utility-preservation argument weakens, because the post-deletion delta is measured against an already-degraded baseline.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Organizations deploying ML-based security tools could respond to GDPR/CCPA data deletion requests in seconds rather than minutes, without taking their detection systems offline for full retraining.
  • The shard-count sweet spot (M=5–10) provides a concrete deployment guideline: enough shards to localize retraining cost, but few enough that each shard retains sufficient data diversity for stable policy learning.
  • The Q-margin (the difference between Q-values for the two actions) serves as a natural confidence score for both ROC analysis and membership-inference auditing, suggesting that value-based RL methods have a built-in privacy audit signal that policy-gradient methods lack.
  • The collapse of the M=20 configuration under aggressive deletion reveals a fundamental data-budget constraint: shard-based unlearning systems must ensure that no shard's retained data falls below a minimum viability threshold, which scales with both shard count and expected cumulative deletion volume.

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 / 7 minor

Summary. This manuscript proposes integrating a Double Deep Q-Network (DDQN) ransomware detector with multi-shard SISA (Sharded, Isolated, Sliced, Aggregated) retraining to enable privacy-compliant machine unlearning under GDPR/CCPA-style deletion requests. The DDQN learns a reward-guided detection policy with asymmetric security costs (false negatives penalized more than false positives), while SISA localizes retraining to affected shards. The framework is evaluated across 270 configurations (shard counts M=5,10,20; forget fractions 1%,5%,10%; deletion rounds 1,5,10; single- and multi-shard scenarios) on a balanced 2,000-sample Windows 11 behavioral dataset with 103 features. Four evaluation dimensions are assessed: utility preservation, oracle-verified forgetting, membership-inference privacy risk via Q-margin ROC-AUC, and computational efficiency. The baseline DDQN achieves F1=0.9925 and AUC=0.9983. The authors find that moderate shard counts (M=5-10) provide the best efficiency-stability trade-off, reducing retraining time to 5-30s versus 80-330s for full retraining, while maintaining MIA scores near 0.5. The paper honestly reports failure modes, including ensemble collapse for M=20 under aggressive repeated deletions.

Significance. The integration of value-based deep reinforcement learning with SISA-based machine unlearning is a novel and practically motivated contribution, as prior SISA work has focused on supervised learning and the combination with RL-based cybersecurity detection is largely unexplored. The dual-audit protocol combining oracle-verified forgetting and Q-margin-based membership-inference auditing is a thoughtful evaluation design that goes beyond runtime-only metrics common in unlearning papers. The systematic deletion-stress evaluation (270 configurations with 5-fold CV) and transparent reporting of boundary-condition failures (M=20 collapse) are commendable. The practical relevance to compliance-oriented security operations is clear. However, the significance is tempered by the small dataset size (2,000 samples), the use of a lightweight MIA proxy rather than adaptive attacks, and an unresolved concern about retraining budget adequacy that potentially affects the privacy claims.

major comments (4)
  1. §3.5.3, Table 4: The unlearning retraining budget is set to 2,000 steps per affected shard, while initial SISA training uses 5,000 steps per shard and oracle retraining uses 5,000 steps. This 60% reduction for unlearning retraining is not justified, and no convergence diagnostic (loss curves, Q-value distributions, reward stabilization) is reported for the retrained shards. This is load-bearing for the privacy claim: the MIA proxy (Eqs. 24-25) distinguishes members from non-members using ensemble Q-margin ROC-AUC. If 2,000 steps is insufficient for the retrained DDQN to converge, Q-margins would be compressed toward zero, artificially pushing MIA AUC toward 0.5 regardless of whether genuine unlearning occurred. The favorable privacy result (MIA ≈ 0.5 across most configurations in Table 9) could thus be an artifact of undertraining rather than evidence of successful forgetting. The asymm_
  2. §5.1, Table 5 vs. Tables 6-7: The baseline (non-SISA) DDQN achieves F1=0.9925, but the SISA ensemble's pre-unlearning F1 scores are notably lower: 0.97968 (M=5), 0.96802 (M=10), and 0.95257 (M=20) as shown in the 'F1 Before' columns of Tables 6-7. This represents a pre-unlearning degradation of 1.3-4.0 percentage points that is not analyzed in depth. The paper focuses on post-unlearning delta (ΔF1) but does not discuss why the SISA ensemble underperforms the global model before any deletion occurs. This matters because the central claim of 'high detection performance' is typically benchmarked against the non-SISA baseline. The authors should explicitly acknowledge and explain this pre-unlearning gap (e.g., reduced per-shard sample diversity, ensemble aggregation effects) and clarify whether the reported F1>0.99 claim in the abstract refers only to the non-SISA baseline.
  3. §3.6.3, Eqs. 24-25: The membership inference 'attack' is a proxy based on Q-margin ROC-AUC between training and test samples. This is a lossless/loss-threshold MIA at best and does not represent an adaptive adversary. The paper states in §6 that 'relied on a Q-margin-based MIA proxy rather than stronger adaptive attacks' is a limitation, but the abstract and conclusions present MIA ≈ 0.5 as evidence of 'limited privacy leakage.' The gap between the proxy metric and actual adversarial privacy risk should be more clearly scoped: the current evaluation shows that Q-margin distributions do not distinguish members from non-members, but this does not rule out leakage detectable by more sophisticated attacks (e.g., neural-network-based MIA, calibration attacks). The claims in the abstract and §5.4 should be qualified accordingly.
  4. §3.1, §5: The dataset comprises 2,000 samples (1,000 ransomware, 1,000 benign) with 103 features. While the paper acknowledges this as a limitation in §6, several claims depend on scale. The computational efficiency claim (5-30s vs. 80-330s for full retraining) is measured on a dataset where full retraining takes only ~34s for a single DDQN (Table 5). The 80-330s oracle times in Table 10 likely reflect 5-fold CV overhead. The scalability of SISA's efficiency advantage to larger datasets (where full retraining would take hours, not seconds) is not demonstrated. The paper should clarify whether the efficiency results are expected to hold proportionally at scale or whether the overhead of managing M independent shard models introduces fixed costs that would change the trade-off at larger N.
minor comments (7)
  1. §3.3, Eq. 9: The reward function uses values +1, -2, -0.5. These are listed as free parameters in the axiom ledger but their sensitivity is not analyzed. A brief note on why these specific ratios were chosen (e.g., 4:1 FN:FP ratio) would strengthen reproducibility.
  2. Table 1: The family name 'GandCrab (Grancrab)' appears inconsistent. The parenthetical 'Grancrab' should be clarified or corrected.
  3. §5.2: The explanation for M=20 collapse under 10%×10 rounds is well-reasoned (shard starvation below 10-sample minimum), but the minimum retained shard size threshold of 10 (Table 4) is introduced without prior justification. Why 10 and not 20 or 50?
  4. Figure 6: The axis labels and configuration codes (e.g., 'S | M10-F5-R5') are compact but require careful parsing. A legend or more descriptive labels would improve readability.
  5. §3.5.4, Eq. 18: Majority voting is used for aggregation, but the paper does not discuss ties (e.g., M=2 or even M with split votes). While M∈{5,10,20} avoids exact ties, the aggregation method should note this assumption.
  6. Table 11: The comparison with prior work is contextual (different datasets/protocols), which the authors acknowledge. However, adding a column for dataset size or evaluation protocol would help readers assess comparability.
  7. §3.3: The discount factor γ=0.1 is described as emphasizing immediate rewards. This is reasonable for per-sample classification, but the paper could note that this effectively reduces the MDP to a contextual bandit, which has implications for the 'reinforcement learning' framing.

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper's derivation chain is self-contained and does not exhibit circularity. The core components—DDQN, SISA, and the MIA proxy—are drawn from external literature (Bourtoule et al. 2021; Hasselt et al. for DDQN). The oracle model (Eq. 23) is trained from scratch on retained data, providing an independent ground truth for verification rather than being defined in terms of the unlearned model. The Q-margin (Eq. 19) is derived from the DDQN's learned Q-values, not fitted to the evaluation metric. The MIA proxy (Eq. 24-25) uses these margins to measure distinguishability, which is an independent measurement, not a self-definitional loop. While there is a self-citation to a preliminary version (Ferdous et al. 2026), it is explicitly described as a limited one-shard extension and does not serve as a load-bearing premise for the present paper's central claims. The skeptic's concern about undertraining (2,000 vs. 5,000 steps) is a validity threat regarding convergence verification, not a circularity issue, as the evaluation metrics are not defined in terms of their own inputs by construction.

Axiom & Free-Parameter Ledger

7 free parameters · 4 axioms · 0 invented entities

The paper introduces no new theoretical entities, particles, or mathematical objects. All components (DDQN, SISA, Q-margin, MIA) are drawn from prior literature. The free parameters are standard ML hyperparameters plus task-specific reward weights. The axioms are domain assumptions inherited from the SISA and RL literature, not ad hoc constructions specific to this paper.

free parameters (7)
  • Reward weights (+1, -2, -0.5) = +1/-2/-0.5
    Asymmetric reward for correct/FN/FP, chosen to encode security cost asymmetry. Not tuned systematically; stated as domain-appropriate without sensitivity analysis.
  • Discount factor gamma = 0.1
    Set low to emphasize immediate rewards. Justified by task structure (independent classifications) but no sensitivity analysis provided.
  • Learning rate alpha = 0.001
    Standard Adam optimizer default; not tuned.
  • Epsilon decay schedule = 1.0 to 0.05 over 5000 steps
    Standard epsilon-greedy schedule; not justified against alternatives.
  • Shard counts M = 5, 10, 20
    Three granularity levels chosen to span coarse-to-fine; reasonable but not exhaustive.
  • Unlearning retraining budget = 2000 steps
    Fixed at 2000 steps per affected shard; no justification for why this is sufficient vs. the 5000-step initial training budget.
  • Minimum retained shard size = 10 samples
    Threshold below which shard models are excluded from ensemble; introduced ad hoc to handle shard exhaustion.
axioms (4)
  • domain assumption SISA shard isolation localizes sample influence to its shard
    Inherited from Bourtoule et al. 2021; foundational to the entire framework. Invoked in Section 3.5.
  • domain assumption DDQN Q-margins serve as valid confidence scores for MIA auditing
    Assumed in Section 3.6.3; the Q-margin difference between training and test samples is used as a membership signal without validating that this proxy correlates with actual attack success.
  • domain assumption Majority voting across shard models preserves detection quality
    Invoked in Section 3.5.4; assumed but the pre-unlearning F1 drop (0.9925 to 0.968-0.979) suggests some degradation.
  • domain assumption Behavioral features from ANY.RUN sandbox are representative of real-world ransomware
    Stated in Section 3.1; sandbox evasion and environment-specific artifacts could limit generalization.

pith-pipeline@v1.1.0-glm · 33593 in / 2698 out tokens · 320994 ms · 2026-07-10T00:03:35.182209+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Auditable Machine Unlearning for Privacy-Compliant Ransomware Detection Using Multi-Shard SISA and Deep Reinforcement Learning." pith.science (2026). https://pith.science/paper/X5SKH6CV

@misc{pith2026260706860,
  author       = {Pith},
  title        = {Pith review of: Auditable Machine Unlearning for Privacy-Compliant Ransomware Detection Using Multi-Shard SISA and Deep Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X5SKH6CV}},
  note         = {Machine review of arXiv:2607.06860}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Ransomware poses an escalating cybersecurity threat as attackers continuously modify behavioral patterns to evade static defenses. Although existing machine learning-based detectors often achieve strong predictive performance, they generally assume fixed training data and do not support the selective removal of previously learned samples. This limitation conflicts with privacy regulations such as the GDPR and CCPA, which require the removal of sensitive user data upon request. To address this challenge, we propose an auditable ransomware detection and unlearning framework that integrates deep reinforcement learning with multi-shard SISA retraining. In the proposed system, a Double Deep Q-Network (DDQN) learns a reward-guided detection policy from behavioral features under asymmetric security costs, while multi-shard SISA enables privacy-compliant selective sample removal through shard-level retraining. The framework was evaluated using four criteria: utility preservation, oracle-based forgetting validation, membership inference auditing, and computational efficiency. On a balanced Windows 11 behavioral dataset comprising 2,000 samples and 103 features, the baseline DDQN detector achieved an F1 score of 0.9925 and an AUC of 0.9983. The experimental results show that single-shard unlearning maintains minimal utility degradation and low oracle disagreement, whereas moderate shard counts (M = 5-10) provide the best efficiency-performance trade-off, reducing retraining time to 5-30 s compared with 80-330 s for full retraining. In addition, the membership inference scores remain close to 0.5 across most configurations, indicating limited privacy leakage after unlearning. These findings demonstrate that a privacy-compliant ransomware detection framework can jointly achieve high detection performance, auditable deletion verification, and efficient sample removal.

Figures

Figures reproduced from arXiv: 2607.06860 by Jannatul Ferdous, Md Zahidul Islam, Rafiqul Islam.

Figure 1
Figure 1. Figure 1: Proposed DDQN–SISA machine unlearning workflow. The framework integrates behavioral dataset construction, RL￾based baseline model selection, multi-shard SISA retraining, and audit-oriented evaluation for privacy-compliant ransomware detection. computational efficiency. Together, these components form a unified system in which detection accuracy, deletion fi￾delity, privacy auditing, and efficiency are join… view at source ↗
Figure 2
Figure 2. Figure 2: Multi-shard SISA framework for DDQN-based machine unlearning, showing four phases: stratified sharding, isolated shard-wise DDQN training, localized deletion-driven retraining of affected shards only, and majority-vote aggregation with ensemble Q-margin scoring. : Preprint submitted to Elsevier Page 7 of 19 [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: illustrates strong class separability, while inference la￾tency remained negligible for both models; DDQN incurred only a modest increase in training and inference time [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Heatmap of utility drop after machine unlearning across forget fractions (1%, 5%, and 10%), deletion rounds (R1, R5, and R10), and SISA configurations (single- and multi-shard with 𝑀 = {5, 10, 20}). Darker colors indicate lower performance degradation, whereas brighter colors indicate greater utility loss [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Oracle disagreement (mean ± std over five folds) under single-shard (a) and multi-shard (b) retraining for shard counts 𝑀 = {5, 10, 20}, showing forget-set (top row) and retained-set (bottom row) disagreements across forget fractions (1%, 5%, 10%) and deletion rounds (𝑅 = 1, 5, 10). substantially more efficient than full retraining, and moder￾ate shard counts (𝑀 = 5 –10) provide the most favorable balance … view at source ↗
Figure 6
Figure 6. Figure 6: Membership-inference privacy change (Δ𝑀𝐼𝐴) across single- and multi-shard unlearning configurations under varying shard counts, forget fractions, and deletion rounds. Values near zero indicate negligible change in attack success after unlearning, whereas positive deviations under extreme deletion settings indicate modest increases in membership distinguishability. : Preprint submitted to Elsevier Page 15 o… view at source ↗
Figure 7
Figure 7. Figure 7: Contour visualization of unlearning time across shard granularities (𝑀) and deletion rounds (𝑅) for different forget fractions (𝑓). The top and bottom rows show single- and multi-shard retraining. The results indicate that single-shard retraining maintains a stable cost, whereas multi-shard retraining is more sensitive to deletion intensity. : Preprint submitted to Elsevier Page 16 of 19 [PITH_FULL_IMAGE:… 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

32 extracted references · 32 canonical work pages · 1 internal anchor

  1. [1]

    URL:https://www.trendmicro

    Trend Micro. URL:https://www.trendmicro. com/vinfo/au/security/news/cybercrime-and-digital-threats/ narrowed-sights-bigger-payoffs-ransomware-in-2019. accessed: Oct. 01,

    work page 2019

  2. [2]

    Behavior-based ransomware classification: A particle swarm optimization wrapper- based approach for feature selection

    Abbasi,M.S.,Al-Sahaf,H.,Mansoori,M.,Welch,I.,2022. Behavior-based ransomware classification: A particle swarm optimization wrapper- based approach for feature selection. Applied Soft Computing 121, 108744. URL:https://www.sciencedirect.com/science/article/pii/ S1568494622001867, doi:https://doi.org/10.1016/j.asoc.2022.108744. abuse.ch,

    work page doi:10.1016/j.asoc.2022.108744 2022

  3. [3]

    IEEE Access 13, 53159–53170

    Xrguard: A model-agnostic approach to ransomware detection using dynamic analysis and explainable ai. IEEE Access 13, 53159–53170. doi:10.1109/ACCESS.2025.3553562. ANY.RUN,

    work page doi:10.1109/access.2025.3553562 2025

  4. [4]

    Gad-sisa: A scalable defense against label flipping attack, in: 2025 IEEE 17th International Conference on Computational Intelligence and Communication Networks (CICN), pp. 893–898. doi:10.1109/CICN67655.2025.11368228. Bourtoule,L.,Chandrasekaran,V.,Choquette-Choo,C.A.,Jia,H.,Travers, A., Zhang, B., Lie, D., Papernot, N.,

    work page doi:10.1109/cicn67655.2025.11368228 2025

  5. [5]

    Machine unlearning, in: 2021 IEEE symposium on security and privacy (SP), pp. 141–159. doi:10.1109/SP40001.2021.00019. Braue, D.,

    work page doi:10.1109/sp40001.2021.00019 2021

  6. [6]

    Cyberse- curity Ventures

    Global ransomware damage costs.https:// cybersecurityventures.com/ransomware-market-report-2022. Cyberse- curity Ventures. Accessed: Feb. 11,

    work page 2022

  7. [7]

    Computers & Security 150, 104293

    Ransoguard: A rnn-based framework leveraging pre-attack sensitive apis for early ransomware detection. Computers & Security 150, 104293. URL:https: //www.sciencedirect.com/science/article/pii/S0167404824005996, doi:https://doi.org/10.1016/j.cose.2024.104293. Cui, B., Hu, Y., Qu, T., He, Y., Sun, L.,

    work page doi:10.1016/j.cose.2024.104293 2024

  8. [8]

    High-Confidence Computing 6, 100338

    A novel zero- day ransomware detection approach based on cvae and 1d- cnn. High-Confidence Computing 6, 100338. URL:https: //www.sciencedirect.com/science/article/pii/S266729522500042X, doi:https://doi.org/10.1016/j.hcc.2025.100338. Deng, X., Cen, M., Jiang, M., Lu, M.,

    work page doi:10.1016/j.hcc.2025.100338 2025

  9. [9]

    Privacy-Aware Machine Unlearning with SISA for Reinforcement Learning-Based Ransomware Detection

    Privacy-aware machine unlearn- ing with sisa for reinforcement learning-based ransomware detection. arXiv preprint arXiv:2604.16760 doi:https://doi.org/10.48550/arXiv. 2604.16760. Ferdous,J.,Islam,R.,Mahboubi,A.,Islam,M.Z.,2025. Anoveltechnique for ransomware detection using image based dynamic features and transfer learning to address dataset limitations...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv 2025

  10. [10]

    IEEE Access 12, 136666–136695

    Ai-based ransomware detection: A comprehensive review. IEEE Access 12, 136666–136695. doi:10.1109/ACCESS.2024.3461965. Gulmez, S., Gorgulu Kakisim, A., Sogukpinar, I.,

    work page doi:10.1109/access.2024.3461965 2024

  11. [11]

    Computers & Security 139, 103703

    Xran: Explainable deep learning-based ransomware detection using dynamic analysis. Computers & Security 139, 103703. URL:https: //www.sciencedirect.com/science/article/pii/S016740482400004X, doi:https://doi.org/10.1016/j.cose.2024.103703. :Preprint submitted to Elsevier Page 18 of 19 Hampton, N., Baig, Z., Zeadally, S.,

    work page doi:10.1016/j.cose.2024.103703 2024

  12. [12]

    Journal of Information Security and Applications 40, 44–51

    Ransomware behavioural analysis on windows platforms. Journal of Information Security and Applications 40, 44–51. doi:10.1016/j.jisa.2018.02.008. Herrera-Silva,J.A.,Hernández-Álvarez,M.,2023. Dynamicfeaturedataset for ransomware detection using machine learning algorithms. Sen- sors

    work page doi:10.1016/j.jisa.2018.02.008 2018

  13. [13]

    Computers & Security 135, 103510

    Behavioral fingerprinting to detect ran- somware in resource-constrained devices. Computers & Security 135, 103510. URL:https://www.sciencedirect.com/science/article/pii/ S0167404823004200, doi:https://doi.org/10.1016/j.cose.2023.103510. Iosifache,

    work page doi:10.1016/j.cose.2023.103510 2023

  14. [14]

    JournalofNetworkandComputerApplications230,103950

    Leveraging appli- cationpermissionsandnetworktrafficattributesforandroidransomware detection. JournalofNetworkandComputerApplications230,103950. doi:10.1016/j.jnca.2024.103950. Jiang, Y., Yu, G., Wang, Q., Wang, X., Ma, B., Sun, C., Ni, W., Liu, R.P.,

    work page doi:10.1016/j.jnca.2024.103950 2024

  15. [15]

    Split unlearning, in: Proceedings of the 2025 ACM SIGSAC ConferenceonComputerandCommunicationsSecurity,Associationfor Computing Machinery, New York, NY, USA. p. 948–962. URL:https: //doi.org/10.1145/3719027.3744787, doi:10.1145/3719027.3744787. Kumar, S., Singh, S.K., Sarin, S., Subba, C.K., Arya, V., Nand- hini, N., Gupta, B.B., Chui, K.T.,

    work page doi:10.1145/3719027.3744787 2025

  16. [16]

    Cyber Security and Applications 3, 100095

    Leveraging dynamic embeddings and reinforcement learning with bayesian networks for ransomware resiliences. Cyber Security and Applications 3, 100095. URL:https://www.sciencedirect.com/science/article/pii/ S2772918425000128, doi:https://doi.org/10.1016/j.csa.2025.100095. Li, N., Zhou, C., Gao, Y., Chen, H., Zhang, Z., Kuang, B., Fu, A.,

    work page doi:10.1016/j.csa.2025.100095 2025

  17. [17]

    IEEE Transactions on Neural Networks and Learning Sys- tems 36, 13709–13729

    Machine unlearning: Taxonomy, metrics, applications, challenges, and prospects. IEEE Transactions on Neural Networks and Learning Sys- tems 36, 13709–13729. doi:10.1109/TNNLS.2025.3530988. Liu,Z.,Ye,H.,Chen,C.,Zheng,Y.,Lam,K.Y.,2025. Threats,attacks,and defenses in machine unlearning: A survey. IEEE Open Journal of the Computer Society 6, 413–425. doi:10....

    work page doi:10.1109/tnnls.2025.3530988 2025

  18. [18]

    Computers & Security 130, 103265

    Improv- ing ransomware detection based on portable executable header using xception convolutional neural network. Computers & Security 130, 103265. URL:https://www.sciencedirect.com/science/article/pii/ S016740482300175X, doi:10.1016/j.cose.2023.103265. Nguyen, T.T., Huynh, T.T., Ren, Z., Nguyen, P.L., Liew, A.W.C., Yin, H., Nguyen, Q.V.H.,

    work page doi:10.1016/j.cose.2023.103265 2023

  19. [19]

    ACM Transactions on Intelligent Systems and Technology 16, 1–46

    A survey of machine unlearning. ACM Transactions on Intelligent Systems and Technology 16, 1–46. URL: https://doi.org/10.1145/3749987, doi:10.1145/3749987. Palo Alto Networks,

    work page doi:10.1145/3749987

  20. [20]

    Romdhana, A., Merlo, A., Ceccato, M., Tonella, P.,

    Adverl-elf: Adversarial elf malware generation using reinforcement learning , 1–7doi:10.1109/ DAC63849.2025.11132466. Romdhana, A., Merlo, A., Ceccato, M., Tonella, P.,

    work page arXiv 2025

  21. [21]

    Computers & Security 131, 103311

    Assessing the security of inter-app communications in android through reinforcement learning. Computers & Security 131, 103311. URL:https: //www.sciencedirect.com/science/article/pii/S0167404823002213, doi:https://doi.org/10.1016/j.cose.2023.103311. Rookard, C., Khojandi, A.,

    work page doi:10.1016/j.cose.2023.103311 2023

  22. [22]

    Computers & Security 140, 103786

    Rriot: Recurrent reinforcement learning for cyber threat detection on iot devices. Computers & Security 140, 103786. URL:https://www.sciencedirect.com/science/article/pii/ S0167404824000877, doi:https://doi.org/10.1016/j.cose.2024.103786. Sehgal, S., Bansal, H., Verma, A.,

    work page doi:10.1016/j.cose.2024.103786 2024

  23. [23]

    Extending grapheraser: Scalable and secure partition-based graph unlearning on large-scale network, in: 2025 IEEE 5th International Conference on ICT in Business Industry & Government (ICTBIG), pp. 1–9. doi:10.1109/ICTBIG68706.2025. 11323936. SnapFiles,

    work page doi:10.1109/ictbig68706.2025 2025

  24. [24]

    Accessed: Sep

    Maturing criminal marketplaces present new challenges to defenders (sophos 2023 threat report).https://www.scribd.com/ document/628559505/Sophos-Threat-Report-2023. Accessed: Sep. 21,

    work page arXiv 2023

  25. [25]

    Ransomware Threat Landscape

    The ransomware threat landscape. Ransomware Threat Landscape. doi:10.2307/j.ctv1f8xc7v. Tang, Y., Sun, J., Wang, H., Deng, J., Tong, L., Xu, W.,

    work page doi:10.2307/j.ctv1f8xc7v

  26. [26]

    Fleet sizing and pricing for hybrid ownership of shared autonomous vehicles in a multimodal transportation system , journal =

    A method of network attack-defense game and collaborative defense decision-making based on hierarchical multi-agent reinforcement learning. Computers & Security 142, 103871. URL:https: //www.sciencedirect.com/science/article/pii/S016740482400172X, doi:https://doi.org/10.1016/j.cose.2024.103871. ThorstenSick,I.P.,2024. Whatiscuckoosandbox?https://cuckoosan...

    work page doi:10.1016/j.cose.2024.103871 2024

  27. [27]

    Available online:https://www.trendmicro.com/ vinfo/gb/security/news/ransomware-by-the-numberst

    Phobos emerges as a formidable threat in q1 2024, lockbit stays in the top spot. Available online:https://www.trendmicro.com/ vinfo/gb/security/news/ransomware-by-the-numberst. VirusShare,

    work page 2024

  28. [28]

    Expert Systems with Applications 296, 129225

    A swarm intelligence-enhanced ransomware detection framework via dynamic memory feature optimization. Expert Systems with Applications 296, 129225. URL:https://www.sciencedirect.com/science/article/pii/ S0957417425028416, doi:https://doi.org/10.1016/j.eswa.2025.129225. Xiao, F., Chen, S., Yang, J., He, H., Jiang, X., Tan, X., Jin, D.,

    work page doi:10.1016/j.eswa.2025.129225 2025

  29. [29]

    Computers & Security 148, 104180

    Grain: Graph neural network and reinforcement learning aided causality discovery for multi-step attack scenario reconstruction. Computers & Security 148, 104180. URL:https: //www.sciencedirect.com/science/article/pii/S0167404824004851, doi:https://doi.org/10.1016/j.cose.2024.104180. Xu,J.,Wu,Z.,Wang,C.,Jia,X.,2024. Machineunlearning:Solutionsand challenge...

    work page doi:10.1016/j.cose.2024.104180 2024

  30. [30]

    Mobtext-sisa: Efficient machine unlearning for mobility logs with spatio-temporal and natural-language data, in: Proceedings of the 33rd ACMInternationalConferenceonAdvancesinGeographicInformation Systems, Association for Computing Machinery, New York, NY, USA. p. 1186–1189. URL:https://doi.org/10.1145/3748636.3763226, doi:10. 1145/3748636.3763226. Zscaler,

    work page doi:10.1145/3748636.3763226

  31. [31]

    Accessed: Oct

    2022 threatlabz state of ransomware report.https://www.zscaler.com/resources/industry-reports/ 2022-threatlabz-ransomware-report.pdf. Accessed: Oct. 12,

    work page 2022

  32. [32]

    Enhancingdynamicmalwarebehavioranalysis throughnovelwindowseventswithmachinelearning

    Önal,G.,Güven,M.,2025. Enhancingdynamicmalwarebehavioranalysis throughnovelwindowseventswithmachinelearning. IEEEAccess13, 153937–153958. doi:10.1109/ACCESS.2025.3604979. :Preprint submitted to Elsevier Page 19 of 19

    work page doi:10.1109/access.2025.3604979 2025