Related-work addendum (Phase 7)¶
Written 15 Oct 2026 by Claude Code for the researcher to check. It covers the “newer neighbours” of Appendix G of the implementation plan (docs/source/DriftWatch_Implementation_Plan.pdf, pp. 50 to 51), plus the items that Appendix G lists as “not re-read”. Every entry was read at the abstract level only, from the arXiv abstract page, on 2026-10-15; none of the full papers was read, so a column entry of “not in the abstract” means exactly that and is not a statement about the paper. Please read each paper before citing it. Descriptions are my paraphrase of the abstract.
The guide’s own related-work table (section 4) stands and is not repeated here.
How to read the table¶
Three axes of this study, as in the guide:
Delay: does the work treat the time a requested label takes to arrive as a parameter?
Budget: does it fix or study the number of labels an analyst can buy?
Conformal or monitor: does it use conformal p-values, or a drift monitor, as a signal?
“Not in the abstract” is not “no”.
Neighbours¶
Work (arXiv, date) |
What the abstract says |
Delay |
Budget |
Conformal or monitor |
Position against this study |
|---|---|---|---|---|---|
Minnei, Manca, Piras, Sotgiu, Pintor, Ghiani, Maiorca, Giacinto, Biggio. Label-efficient Training Updates for Malware Detection over Time, 2603.28396 (30 Mar 2026) |
A model-agnostic framework that tests active and semi-supervised learning, alone and combined, on Android and Windows malware; reports annotation cost cut by up to 90% at detection performance near full-label retraining; adds a feature-level drift analysis that relates feature stability to detector performance |
Not in the abstract |
Central: annotation cost is the objective, though a fixed per-slot budget is not described |
Feature-level drift analysis; no conformal in the abstract |
The closest neighbour on budget and on the two platforms. This study fixes the budget per slot, makes label arrival a parameter, and reports A-AUT. Compare after reading how its time slices and its labelling cost are defined |
Herzog, Eusebi, Cavallaro. On the Reliability and Stability of Selective Methods in Malware Classification Tasks (Aurora), 2505.22843 (28 May 2025, revised 21 Jan 2026) |
Asks whether drift-adaptive malware classifiers hold up in operation beyond standard metrics; Aurora checks whether confidence estimates are trustworthy and performance is stable over time; finds current approaches fragile across datasets with different drift |
Not in the abstract |
Motivation: unreliable confidence wastes annotation budgets in active learning |
Confidence-error alignment and selective classification; no conformal in the abstract |
Relevant to RQ3: selection by confidence is only as good as the confidence. Cite as a caution about the credibility and uncertainty selectors |
Kalny, Jurecek, Stamp. Detecting Concept Drift in Evolving Malware Families Using Rule-Based Classifier Representations, 2604.22629 (24 Apr 2026) |
Structural drift detection from rule sets of decision trees trained across temporal windows on EMBER2024; fixed two-month windows with feature-level Pearson correlation worked most reliably; no single method dominates; baselines include RIPPER and Transcendent |
Not in the abstract |
Not in the abstract |
Drift monitoring (structure-based) |
A monitor-side neighbour for RQ4 and for the EMBER2024 tier. Its window finding is relevant to the slot-length question (ADR-0004) |
Li and Bertino. LFreeDA: Label-Free Drift Adaptation for Windows Malware Detection, 2511.14963 (18 Nov 2025) |
Label-free adaptation: unsupervised domain adaptation on malware images generates and filters pseudo-labels, then a control-flow-graph classifier is adapted; reports up to 12.6% accuracy and 11.1% F1 gains over no adaptation, near methods given 300 true labels |
Not in the abstract |
Label-free by design |
Avoids drift detection by design |
The zero-label end of the budget axis, for Windows. A reference point for “what a budget of zero buys” if a pseudo-label arm is ever added; not used in this study |
Alam, Piplai, Rastogi. ADAPT: A Pseudo-labeling Approach to Combat Concept Drift in Malware Detection, 2507.08597 (11 Jul 2025, revised 3 Aug 2025) |
Model-agnostic semi-supervised pseudo-labelling to use unlabeled data under drift; works with neural and tree models; five datasets across Android, Windows and PDF |
Not in the abstract |
Only indirectly (avoids ground-truth labels) |
Not in the abstract |
Same family as LFreeDA: adaptation without true labels. Context for the statement that true labels are the scarce resource |
Zheng, Yang, Ngai, Jana, Cavallaro. TIF: Learning Temporal Invariance in Android Malware Detectors, 2502.05098 (7 Feb 2025, revised 22 Jun 2026) |
A training framework that groups data into environments by observation date and uses contrastive learning with invariant gradient alignment so that detectors generalise across time; reports gains especially in early deployment |
Not in the abstract |
Not in the abstract |
Not in the abstract |
A way to slow decay by training rather than by labelling. Orthogonal to the budget and delay axes; a possible future model arm |
Roh, Kaya, Kruegel, Vigna, Hong. MADCAT: Combating Malware Detection Under Concept Drift with Test-Time Adaptation, 2505.18734 (24 May 2025) |
Self-supervised encoder-decoder; adapts at test time using a small balanced sample of test data; outperforms baselines on continuous Android malware detection |
Not in the abstract |
A small test-time sample is used; no labelling budget is described |
Not in the abstract |
Adapts at test time without true labels. Relevant to the question of what “retrain” means in the policy axis |
He, Lei, Qin, Ren, Chen. Combating Concept Drift with Explanatory Detection and Adaptation for Android Malware Classification (DREAM), 2405.04095 (7 May 2024, revised 24 May 2025) |
Drift detection with explanations in a contrastive autoencoder; supports concept revision for retraining; matches the best existing method on a Drebin classifier with 76.6% fewer newly labelled samples |
Not in the abstract |
Reduces expert effort and newly labelled samples; no fixed budget |
Drift detection is central |
A drift-monitor neighbour that also counts labels. Place beside the monitors of RQ4 |
Kamol, Lopez, Nowmi, Rivas, Haque, Raff, Piplai, Rahman. McNdroid: A Longitudinal Multimodal Benchmark for Robust Drift Detection in Android Malware, 2605.06894 (7 May 2026) |
A 2013 to 2025 (without 2015) Android benchmark with static, dynamic and call-graph features; degradation over growing train-test gaps; multimodal fusion beats the best single modality over long gaps |
Not in the abstract |
Not in the abstract |
Drift analysis and benchmarking; a monitor is not described in the abstract |
A dataset neighbour of LAMDA (the same years and gap). Would be a third Android source if licences and size allow; not used |
Yu, Chen, Zhang, Liang, Qiao, Lai. EMA: Efficient Model Adaptation for Learning-based Systems, 2605.13942 (13 May 2026) |
Adapts models in networked systems with state transformers for a warm start and by prioritising which data to label; 14.9% to 42.4% lower adaptation cost across eight systems |
Not in the abstract |
Trade-off between labelling and training cost; no fixed labelling budget |
Not as a monitor |
A systems paper, not malware. Listed in the plan because it prioritises what to label; probably not a neighbour for citation. Kept as a note |
Not identified¶
“Li et al. (NDSS 2025)” appears in Table G.1 of the plan without a title. One web search on 2026-10-15 did not find an NDSS 2025 paper by an author named Li on malware drift. It may be the Li and Bertino paper above (LFreeDA is arXiv only, not NDSS) or a different work. A second search on 2026-10-20 found one candidate: A. S. Li, A. Iyengar, A. Kundu and E. Bertino, “Revisiting Concept Drift in Windows Malware Detection: Adaptation to Real Drifted Malware with Minimal Samples”, NDSS 2025 (arXiv 2407.13918; https://www.ndss-symposium.org/ndss-paper/revisiting-concept-drift-in-windows-malware-detection-adaptation-to-real-drifted-malware-with-minimal-samples/). It argues that active-learning retraining is suboptimal and proposes domain adaptation on control-flow graphs. It fits the context but has not been matched to Table G.1. On 2026-10-22 the paper was confirmed on the NDSS 2025 accepted-papers list (ndss-symposium.org/ndss2025/accepted-papers/) with the same title and the same four authors, and the arXiv record 2407.13918 (v1 18 Jul 2024, updated 19 Dec 2024) gives the same title and authors [L-33]. It is in
paper/references.bibasli2024revisiting. [verify: the researcher confirms that this is the paper Table G.1 means by “Li et al. (NDSS 2025)”.]ActDroid (arXiv 2401.16982), SEED (arXiv 2605.24903), FIRCE (arXiv 2605.01962), arXiv 2606.07789 are in the guide’s table and are covered in
report/ch05_related_work.md; they were not re-read for this addendum.
Position, in the words the guide allows¶
No neighbour read at abstract level combines label delay as a parameter, a fixed analyst budget and conformal drift monitoring, reported with A-AUT on Android and Windows PE data. That means “not found in a limited search”, not “absent”: the search was one session of abstract-level reading and can miss papers whose abstracts do not use these words.
Contribution statement (the guide’s wording): “We study malware detector lifetime under label delay, a fixed analyst budget and conformal drift monitoring in a single time-aware protocol, reported with A-AUT, on Android and Windows PE data.” The word “first” is not used.
What changes in the report¶
report/ch05_related_work.md: add Minnei et al. as the closest budget neighbour, and Aurora next to the selection discussion; list the rest in one sentence each.E3 (Phase 8) discussion: Aurora’s caution about confidence under drift applies to the uncertainty selector and to credibility.
Audit 2026-10-27 (first monthly audit)¶
Run by scripts/arxiv_audit.py for papers published from 2026-09-01 to 2026-10-27 in cs.CR and cs.LG (list in reports/arxiv_audit/2026-10-27.md). 49 candidates matched the keyword filter; the abstracts were read from the arXiv API record on 2026-10-27 (no full paper was read). Five are recorded below, three are adjacent but outside this study and are set aside, and 41 are keyword matches from unrelated fields (biomedicine, robotics, aerospace, language models). “Not in the abstract” is not “no”. Every entry below is marked [verify: read in full before citing].
Work (arXiv, date) |
What the abstract says |
Delay |
Budget |
Conformal or monitor |
Position against this study |
|---|---|---|---|---|---|
HYDRA: Proactive Android Malware Drift Adaptation via Hierarchical Graph Contrastive Learning, 2609.26352 (22 Sep 2026) |
Argues that reactive drift adaptation responds only after performance has dropped and imposes a manual annotation burden; proposes a proactive framework that learns drift-invariant representations from hybrid control-flow and function-call graphs with a cross-domain contrastive objective |
Not in the abstract |
The annotation burden is the motivation; no fixed per-slot budget is described |
Not a monitor; adaptation by representation learning |
Android neighbour on the “adapt without paying for labels” side, like TIF and MADCAT above. Orthogonal to the delay and budget axes; a possible model arm, not a comparison |
Temporal Generalization and Explanation Stability of Control Flow Graph Neural Networks for Malware Detection, 2609.24280 (21 Sep 2026) |
Evaluates graph neural networks on a strict temporal split (train on one period, score once on a later one) over static control-flow graphs of Windows PE files, and finds the message-passing operator changes robustness to the shift |
Not in the abstract |
Not in the abstract |
Not in the abstract |
A Windows PE neighbour for the time-aware evaluation axis (RQ1). It uses one train and one test period, not a rolling A-AUT, and a small corpus; compare after reading its split and its corpus size |
Rethinking Tabular Foundation Models On Data Streams, 2610.05352 (4 Oct 2026) |
A systematic study of tabular foundation models on data streams, covering memory management, cost, concept drift and delayed labels; reports that keeping the most recent examples is as effective as existing memory-management techniques and that these models recover faster than streaming learners after drift |
Yes: delayed labels are among the stream-specific challenges studied |
Memory and cost, not a labelling budget |
Not a monitor |
Not malware, but a recent treatment of delayed labels in streams. Relevant to how the delay axis is framed in chapters 2 and 5; read before citing |
Navigating the Latent Manifold: Proactive Concept Drift Adaptation for Resilient NIDS (DriftXpert), 2609.08623 (8 Sep 2026) |
A two-stage offline framework for network intrusion detection: unsupervised drift detection by outlier analysis in a latent space, then representation-consistency alignment against forgetting |
Not in the abstract |
Not in the abstract |
Drift detection is central (unsupervised) |
A monitor-side neighbour for RQ4, in network intrusion detection rather than malware. Its detector is unsupervised and label-free, like the monitors here |
A Three-Axis Stress Test of LLM vs Classical ML for Network Intrusion Detection under Distribution Shift and Adversarial Evasion, 2609.13511 (11 Sep 2026) |
Compares XGBoost and a language-model classifier on two networks across same-dataset, cross-dataset and evasion tests; finds no universal winner and a large false-positive rate for one model under shift |
Not in the abstract |
Not in the abstract |
Not in the abstract |
Supports the point that a single-dataset score hides a shift effect (RQ1) in a neighbouring domain. Its evasion axis is outside this study and is not used |
Set aside after reading the abstract: AURA: Adaptive Uncertainty-Routed Analysis for Email Threat Detection, 2609.19873 (email, uncertainty routing, not a label or monitoring question); DistillGuard, 2609.28996 (malicious NPM packages; concept drift is named only as a motivation); and Conformal Prediction for Offensive Security, 2609.05165, which concerns using conformal prediction on the attacker’s side and is outside this study’s scope (the repository describes how detectors decay, not how to defeat them).
Combination check: none of the recorded abstracts combines label delay as a parameter, a fixed analyst budget and conformal drift monitoring for malware. This is “not found in the abstracts of this month’s candidates”, the same limited statement as above.
Chapters that may need a change (not made here): report/ch05_related_work.md (HYDRA and 2609.24280 as Android and Windows PE neighbours; 2610.05352 for delayed labels in streams). The researcher decides.