Pith. sign in

REVIEW 4 major objections 5 minor 47 references

Adaptive Intrusion Detection System using Transformer-Based Neural Networks and Continual Learning Approach with Adversarial Investigation

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A benign-anchored replay buffer lets a transformer-based intrusion detection system learn new attack classes without catastrophic forgetting, recovering near-joint-training accuracy on CICIDS2017 under two evaluation scenarios — and the…

desk verdict Worth taking seriously for the CII scenario and buffer-poisoning analysis, but the CII oracle inconsistency and the broken ASR estimator mean the headline numbers need revision before trust. read the letter →

arxiv 2608.04602 v1 pith:UCF5VD72 submitted 2026-08-05 cs.CR

Azizi Ariffin , Afif Haris , Faiz Zaki , Hazim Hanif , Nor Badrul Anuar This is my paper
classification cs.CR MSC 68T0768M25
keywords intrusiondetectioncontinuallearningcatastrophicforgettingexperiencereplaytabulartransformerbenignanchoringbackdoorattacklabelflipping
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that a network intrusion detection system can absorb new attack classes over time without forgetting the ones it already knows, if retraining interleaves a small class-balanced buffer of past flows, with benign traffic guaranteed a place in every update. On the CICIDS2017 benchmark this design reaches 0.9994 accuracy under the standard class-incremental setup and 0.9989 under the paper's new class-instance incremental setup — where benign flows reappear alongside every new attack — while sequential fine-tuning, EWC, and LwF collapse. The same buffer, however, is the system's weak point: flipping the labels of stored exemplars at a 1% budget drops accuracy from 0.9995 to 0.0053, and a backdoor hidden in packet-timing features keeps overall accuracy above 0.97 while sending 95–100% of triggered attack flows through as benign. The paper concludes that replay with an explicit benign anchor is the best-performing adaptive-IDS configuration it observed, and that buffer integrity is a strict operational requirement.

What carries the argument

The central mechanism is the coupling of two components. First, a tabular transformer encoder that embeds each categorical flow field (protocol, flags, service) into a learned vector and runs multi-head self-attention over the token sequence, so the representation of one field depends on the values of the other fields in the same flow record; the classifier head expands by appending output units when new classes arrive, with existing units copied unchanged. Second, a benign-anchored class-balanced experience replay buffer: at the end of each experience, every class seen so far contributes a per-class budget $b$ (1%, 5%, or 10%) of its training flows via class-balanced reservoir sampling, and benign flows are guaranteed representation in every replay batch. The class-instance incremental scenario is the evaluation mechanism that stresses this design by forcing the benign–new-attack decision boundary to be re-learned at every experience. The adversarial investigation uses two poisoning mechanisms on the buffer: overt label flipping, where every stored exemplar's label is reassigned, and a stealthy clean-label backdoor that adds a fixed perturbation to the packet inter-arrival-time features listed in the paper's Table 2, with the poisoned exemplars labelled benign so the model learns to treat trigger-carrying attack flows as benign.

What would settle it

Run a backdoored detector against live or replayed attack flows whose packet inter-arrival times are paced to match the trigger, after passing them through CICFlowMeter and the paper's standardization; if triggered flows are not classified as benign at the claimed 95–100% rate, the deployment threat is not established. Re-running the backdoor evaluation with a reset evaluation set and a single non-repeated injection per experience would also determine whether the three attack-success-rate cells above 100% are purely estimator artifacts.

Watch

Extended reading notes

Core claim

Stated as the authors would state it: a tabular transformer encoder paired with a benign-anchored, class-balanced experience replay buffer recovers near-joint-training performance in a continually learning intrusion detector, and the standard evaluation scenario used by prior work is unrealistically kind. Under the traditional class-incremental (CI) scenario benign traffic appears only in the first experience; the paper introduces the class-instance incremental (CII) scenario in which benign flows are present in every experience alongside newly arriving attack classes, arguing that this mirrors production exposure. The framework achieves 0.9994 accuracy and 0.9953 macro-F1 under CI and 0.9989 accuracy and 0.9960 macro-F1 under CII at a 10% per-class buffer budget, within 0.0003 of the joint oracle on accuracy, with forgetting and intransigence both below 0.01. Benign anchoring is the load-bearing component: re-exposing the benign class lifts memory-free baselines by 0.74–0.82 accuracy under CII. The flip side, which the paper investigates explicitly, is that the buffer is a persistent attack surface: a 1% label-flip budget collapses accuracy to 0.0053, while a backdoor that perturbs inter-arrival-time features of stored exemplars holds overall accuracy above 0.97 and reaches 95–100% attack success on trigger-carrying flows, a corruption an accuracy dashboard would not flag.

Load-bearing premise

The backdoor results rest on the assumption that an attacker can actually realize the inter-arrival-time trigger in live network traffic by pacing packet transmission, and that the fixed trigger perturbation survives feature standardization and the CICFlowMeter pipeline; the paper specifies neither the perturbation magnitude nor the fraction of stored exemplars injected, and offers no live-traffic demonstration.

Editorial extensions

If this is right

  • A deployed IDS can be updated incrementally with a 1–10% per-class replay budget and land within 0.0003 accuracy of a model retrained on all historical data, removing the need for costly full retraining.
  • The benign-anchored CII scenario should become the default evaluation for continual-learning IDS research, because it exposes that accuracy-only reporting flatters regularisation-based methods whose macro-F1 stays below 0.73 even when accuracy recovers.
  • The replay store must be treated as security-critical state: a 1% label-flip budget disables the detector outright, so integrity protection, buffer auditing, and periodic revalidation of stored labels and features are operational requirements.
  • The backdoor result implies that an attacker who can pace packet transmission could route chosen attack traffic past a detector whose top-line metrics look healthy, so monitoring must include attack-success-rate-style checks rather than aggregate accuracy alone.
  • Because the same collapse-and-recovery pattern holds across MLP, CNN, and transformer backbones, the forgetting failure and its replay remedy are properties of the learning scenario rather than of any particular encoder.

Reading between the lines

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

  • The paper attacks the buffer only under the CI scenario; applying the same two attacks under CII, where benign is re-exposed in every experience, is the natural next experiment, and it is not obvious that the backdoor's stealth survives the stronger benign anchoring.
  • The trigger's realizability in live traffic is the paper's unstated load-bearing assumption; a live-traffic demonstration in which an attacker paces packets to hit the inter-arrival-time pattern would extend the result from benchmark to deployment.
  • Three of the nine backdoor attack-success-rate cells exceed 100%, an artifact of the paper's own estimator that it openly explains; a cleaner protocol (one target class per run, a reset evaluation set, a single injection per experience) would sharpen the 95–100% claim.
  • The benign-anchoring insight transfers beyond IDS: any continual classifier facing a persistent majority class, such as fraud detection or medical screening, could adopt the same guarantee of keeping the reference class in every replay batch.
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

4 major / 5 minor

Summary. The paper proposes an adaptive network intrusion detection framework that couples a tabular transformer encoder with a class-balanced experience replay buffer in which the benign class is anchored in every replay update. It introduces a class-instance incremental (CII) scenario in which benign flows appear in every experience, evaluates the framework on CICIDS2017 under both the standard class-incremental (CI) and the proposed CII scenarios, and compares against naive fine-tuning, EWC, LwF, iCaRL, and two experience-replay baselines. The paper separately investigates two poisoning attacks on the replay buffer: label flipping and a clean-label timing-feature backdoor. The main empirical claims are that replay with a modest per-class budget recovers near-joint-training performance under both scenarios, that benign anchoring is essential for preventing forgetting, that label flipping collapses the model, and that the backdoor preserves high overall accuracy while achieving a high attack success rate on trigger-carrying flows.

Significance. If the central claims hold, the paper makes useful contributions: the CII scenario is a more realistic stress test than the usual CI setup, the adversarial treatment of the replay buffer as a persistent attack surface is timely and underexplored, and the cross-architecture benchmark in Section 5.7 supports the conclusion that the advantage of replay over regularization-based methods is scenario-level rather than encoder-specific. The paper also ships a public code repository and is unusually transparent about the flaws in its own backdoor ASR estimator, which is a genuine strength. The two load-bearing concerns are the unspecified construction of the CII training data, which prevents an apples-to-apples interpretation of the CII oracle comparison, and the single-seed evaluation that underlies the near-joint and best-performing claims.

major comments (4)
  1. [§3.6, Table 4; §5.2, Tables 7–8]
  2. [§3.5, §5.2, Table 7]
  3. [§5.6, Table 12]
  4. [§3.9, Tables 11–12]
minor comments (5)
  1. [Table 6]
  2. [§6, fifth takeaway]
  3. [§9 and §12]
  4. [Table 7]
  5. [References, [43]]

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the central claims are empirical comparisons on an external benchmark; the only self-citation is a non-load-bearing reuse of the authors' own backdoor trigger construction.

full rationale

The paper's main claims—that benign-anchored replay recovers near-joint-training performance and that replay-buffer poisoning can collapse or backdoor the model—are empirical results on the external CICIDS2017 benchmark, measured against seven baselines under two scenarios. No performance number is obtained by fitting to the target claim; the replay budget is swept over {1%, 5%, 10%} and all results are reported, including the less flattering ones. The one notable self-citation is Section 3.9's `following the clean-label IAT-trigger construction of Guo et al. (2025b)`, whose authors overlap with the present paper. That citation supplies the trigger design for the backdoor attack, but the attack is actually implemented, executed, and its attack success rate is recomputed from the released training logs, so the measured outcomes do not reduce to the citation. The differing CII oracle (0.9955) versus CI oracle (0.9997) is an internal-consistency concern about how benign flows are reused in the CII union, but it is not a circular step: the CII replay result is compared with the same scenario's oracle and baselines. No fitted parameter is relabelled as a prediction, and no uniqueness theorem is imported from the authors' prior work to forbid alternatives.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claims rest on empirical evaluation with a small set of design choices. The only freely chosen numbers are the memory budget sweep and the unstated trigger parameters. The domain assumptions about dataset realism, scenario fidelity, trigger realizability, and clean labels are the main burdens the reader must accept.

free parameters (2)
  • per-class memory budget b = 1%, 5%, 10% (swept)
    Design choice swept over three values; all results are reported for each budget, so no single value is fitted to the test set.
  • backdoor trigger magnitude and injected fraction
    The paper specifies the IAT feature set but not the perturbation size or the proportion of stored exemplars poisoned; these parameters are left to the released code and are not reported in the text.
assumptions (4)
  • domain assumption CICIDS2017 is a valid proxy for production network traffic
    The paper acknowledges the dataset has mislabeled flows and a limited 2017-era attack inventory (Section 6), so absolute numbers are treated as upper bounds.
  • ad hoc to paper The CII scenario faithfully represents production benign exposure
    The scenario is proposed by the authors and underpins the evaluation of benign anchoring; no external validation of its realism is given.
  • domain assumption An attacker can realize the IAT trigger on live traffic by pacing packet transmission
    Assumed in Section 3.9; no live-traffic experiment demonstrates that the trigger survives the feature extraction pipeline.
  • domain assumption The defender has clean class labels at training time
    Stated in Section 7 as a strong assumption; if labels are noisy, the benign anchor itself becomes a poisoning vector.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adaptive Intrusion Detection System using Transformer-Based Neural Networks and Continual Learning Approach with Adversarial Investigation." pith.science (2026). https://pith.science/paper/UCF5VD72

@misc{pith2026260804602,
  author       = {Pith},
  title        = {Pith review of: Adaptive Intrusion Detection System using Transformer-Based Neural Networks and Continual Learning Approach with Adversarial Investigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UCF5VD72}},
  note         = {Machine review of arXiv:2608.04602}
}
read the original abstract

Network intrusion detection systems (IDS) trained on fixed traffic snapshots decay silently after deployment as threat distributions shift. Fine-tuning models on new attacks triggers catastrophic forgetting, while retraining from scratch is computationally infeasible. Replay-based continual learning counters this, but existing methods unrealistically confine benign traffic to a single early task and ignore the replay buffer as a potential attack surface. To address this, we present an adaptive IDS framework coupling a tabular transformer encoder with a class balanced experience replay buffer that replays benign traffic at every update to stabilize decision boundaries. We introduce the class-instance incremental (CII) scenario where benign flows reappear alongside new attacks as a more faithful stress test, and probe the buffer with overt label flipping and stealthy backdoor poisoning attacks. On the CICIDS2017 benchmark, our framework achieved 0.9994 accuracy under the traditional class incremental setup and 0.9989 under CII, with negligible forgetting, drastically outperforming sequential fine-tuning (0.0052), EWC (0.0324), LwF (0.0699), and iCaRL (0.8770) baselines. While injecting benign traffic into every experience proves essential for preventing forgetting, the replay buffer introduces critical vulnerabilities. Label-flipping collapses the model entirely (0.0053 accuracy at a 1% budget), and the backdoor maintains 0.97 overall accuracy while driving the attack success rate on trigger flows to 95%, evading standard monitoring. Ultimately, while a modest replay budget recovers near-joint-training performance, ensuring buffer integrity emerges as a strict operational requirement.

Figures

Figures reproduced from arXiv: 2608.04602 by the authors.

Figure 1
Figure 1. Overview of the proposed adaptive IDS framework. 3.5. Training procedure All hyperparameters were selected by grid search on the validation set of the first experience. The network is trained with the AdamW optimiser at learning rate 5İ10*4 and weight decay 1İ10*5, with early stopping on validation loss (patience 10 epochs, up to 100 epochs per experience) and a ReduceLROnPlateau schedule (factor 0.1, patience 5). T… view at source ↗
Figure 2
Figure 2. Tabular transformer encoder architecture. Ariffin et al.: Preprint Page 8 of 23 [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Benign-anchored class-balanced experience replay: buffer composition and batch construction [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Experience distribution under the CI and CII scenarios on CICIDS2017. indicate that the learner matches or marginally exceeds the single-snapshot oracle. Reporting forgetting and intransigence together is essential because a trivially stable model that refuses to learn…
Figure 5
Figure 5. Figure 5: Label-flipping poisoning attack on the replay buffer under the adopted threat model [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Backdoor attack pipeline across two experiences: the buffer is rebuilt by random sampling after each experience, the trigger is injected into the stored inter-arrival-time features with the exemplar labels set to benign, and evaluation splits into clean testing (utilit…
Figure 7
Figure 7. Figure 7: CI scenario: accuracy and macro-F1 by continual-learning method. Dashed lines indicate oracle accuracy (0.9997) and macro-F1 (0.9979). ER-Balanced at b = 10˜ closes within 0.0003 of the oracle on accuracy. under CII. The Na"ive sequential model reaches 0.8700 checkpoin…
Figure 8
Figure 8. Figure 8: CII scenario: accuracy and macro-F1 by continual-learning method. Dashed lines indicate oracle accuracy (0.9955) and macro-F1 (0.9795). ER-Balanced at b = 10˜ matches or marginally exceeds the oracle on both metrics. at every budget, and ER-Balanced at b = 10% ends the…
Figure 9
Figure 9. Figure 9: Backdoor attack on the replay buffer: overall accuracy stays above 0.97 even at p = 10˜ poison rate while macro-F1 degrades only modestly from 0.9957 to 0.8933, so the aggregate view registers little. The attack’s real effect is the near-total attack success rate of […
Figure 10
Figure 10. Figure 10: Cross-architecture benchmark: accuracy by backbone (MLP, SGM-CNN, FT-Transformer) and continual-learning strategy under the CI scenario (a) and CII scenario (b). Benign anchoring under CII lifts non-rehearsal strategies by 0.74–0.82 in accuracy while rehearsal-based s…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 31 canonical work pages

  1. [1]

    F., Khraisat, A., S P, S

    A H, N. F., Khraisat, A., S P, S. I., & Li, G. (2025). Adaptive memory replay for network intrusion detection: Tackling data drift and catastrophic for- getting. Computer Networks, 272, 111712.https: //doi.org/10.1016/j.comnet.2025.111712

    work page arXiv 2025

  2. [2]

    Nguyen, T.-T., & Park, M. (2025). EL-GNN: A Continual-Learning-Based Graph Neural Network for Task-Incremental Intrusion Detection Systems. Electronics, 14(14), 2756.https://doi.org/10. 3390/electronics14142756

    work page 2025

  3. [3]

    K., Reddy Tamma, B., & Channappayya, S

    Amalapuram, S. K., Reddy Tamma, B., & Channappayya, S. S. (2024). SPIDER: A Semi- Supervised Continual Learning-based Network In- trusion Detection System. In Proc. IEEE INFO- COM 2024 (pp. 571–580).https://doi.org/10. 1109/INFOCOM52122.2024.10621428

    arXiv 2024

  4. [4]

    K., Channappayya, S

    Amalapuram, S. K., Channappayya, S. S., & Reddy Tamma, B. (2023). Augmented Memory Replay-based Continual Learning Approaches for Network Intrusion Detection. In Advances in Neural Information Processing Systems 36 (pp. 17156–17169)

    work page 2023

  5. [5]

    P., Poornachandran, P., Al-Nemrat, A., & Venka- traman, S

    Vinayakumar, R., Alazab, M., Soman, K. P., Poornachandran, P., Al-Nemrat, A., & Venka- traman, S. (2019). Deep Learning Approach for Intelligent Intrusion Detection System. IEEE Access, 7, 41525–41550.https://doi.org/10. 1109/ACCESS.2019.2895334

    work page arXiv 2019

  6. [6]

    Moustafa, N., & Slay, J. (2015). UNSW-NB15: A Comprehensive Data Set for Network Intrusion Detection Systems. In Proc. Military Communica- tions and Information Systems Conference (Mil- CIS) 2015.https://doi.org/10.1109/MilCIS. 2015.7348942

    work page arXiv 2015

  7. [7]

    Yin, C., Zhu, Y., Fei, J., & He, X. (2017). A Deep Learning Approach for Intrusion De- tection Using Recurrent Neural Networks. IEEE Access, 5, 21954–21961.https://doi.org/10. 1109/ACCESS.2017.2762418

    work page arXiv 2017

  8. [8]

    N., Phai, V

    Shone, N., Ngoc, T. N., Phai, V. D., & Shi, Q. (2018). A Deep Learning Approach to Network IntrusionDetection.IEEETransactionsonEmerg- ing Topics in Computational Intelligence, 2(1), 41–50.https://doi.org/10.1109/TETCI.2017. 2772792

    work page doi:10.1109/tetci.2017 2018

Show all 47 references
  1. [9]

    Huang, X., Khetan, A., Cvitkovic, M., & Karnin, Z. (2020). TabTransformer: Tabular Data Model- ing Using Contextual Embeddings. arXiv preprint arXiv:2012.06678

    2020 arXiv

  2. [10]

    Gorishniy, Y., Rubachev, I., Khrulkov, V., & Babenko, A. (2021). Revisiting Deep Learn- ing Models for Tabular Data. In Advances in Neural Information Processing Systems 34 (pp. 18932–18943)

    2021

  3. [11]

    Ö., & Pfister, T

    Arik, S. Ö., & Pfister, T. (2021). TabNet: At- tentive Interpretable Tabular Learning. In Proc. AAAI Conf. on Artificial Intelligence, 35(8), 6679–6687

    2021

  4. [12]

    B., & Goldstein, T

    Somepalli, G., Goldblum, M., Schwarzschild, A., Bruss, C. B., & Goldstein, T. (2021). SAINT: Im- proved Neural Networks for Tabular Data via Row Attention and Contrastive Pre-training. arXiv preprint arXiv:2106.01342

    2021 arXiv

  5. [13]

    &Hadsell, R

    Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Ve- ness,J.,Desjardins,G.,Rusu,A.A.,... &Hadsell, R. (2017). Overcoming catastrophic forgetting in neural networks. Proc. National Academy of Sci- ences, 114(13), 3521–3526.https://doi.org/10. 1073/pnas.1611835114

    2017

  6. [14]

    Zenke, F., Poole, B., & Ganguli, S. (2017). Con- tinual Learning Through Synaptic Intelligence. In Proc. 34th International Conference on Machine Learning (pp. 3987–3995)

    2017

  7. [15]

    Li, Z., & Hoiem, D. (2018). Learning Without Forgetting. IEEE Trans. Pattern Analysis and Machine Intelligence, 40(12), 2935–2947.https: //doi.org/10.1109/TPAMI.2017.2773081

    2018

  8. [16]

    Lopez-Paz, D., & Ranzato, M. (2017). Gradient Episodic Memory for Continual Learning. In Ad- vances in Neural Information Processing Systems 30 (pp. 6467–6476)

    2017

  9. [17]

    Chaudhry, A., Ranzato, M., Rohrbach, M., & Elhoseiny, M. (2019a). Efficient Lifelong Learning with A-GEM. In Proc. International Conference on Learning Representations. Ariffin et al.: Preprint Page 21 of 23 Adaptive IDS with Transformers, Continual Learning, and Adversarial I...

    2019

  10. [18]

    A., Kolesnikov, A., Sperl, G., & Lam- pert, C

    Rebuffi, S. A., Kolesnikov, A., Sperl, G., & Lam- pert, C. H. (2017). iCaRL: Incremental Classifier and Representation Learning. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (pp. 2001–2010)

    2017

  11. [19]

    Buzzega, P., Boschini, M., Porrello, A., Abati, D., & Calderara, S. (2020). Dark Experience for General Continual Learning: a Strong, Simple Baseline. In Advances in Neural Information Pro- cessing Systems 33 (pp. 15920–15930)

    2020

  12. [20]

    K., Torr, P

    Chaudhry, A., Rohrbach, M., Elhoseiny, M., Ajanthan, T., Dokania, P. K., Torr, P. H. S., & Ranzato, M. (2019b). On Tiny Episodic Memories in Continual Learning. arXiv preprint arXiv:1902.10486

    2019 arXiv

  13. [21]

    (2019).OnlineContinualLearningwithMaximally Interfered Retrieval

    Aljundi, R., Belilovsky, E., Tuytelaars, T., Char- lin, L., Caccia, M., Lin, M., & Page-Caccia, L. (2019).OnlineContinualLearningwithMaximally Interfered Retrieval. In Advances in Neural Infor- mation Processing Systems 32 (pp. 11849–11860)

    2019

  14. [22]

    Chrysakis, A., & Moens, M.-F. (2020). Online Continual Learning from Imbalanced Data. In Proc. 37th International Conference on Machine Learning (pp. 1952–1961)

    2020

  15. [23]

    Riemer, M., Cases, I., Ajemian, R., Liu, M., Rish, I., Tu, Y., & Tesauro, G. (2019). Learning to Learn Without Forgetting by Maximizing Transfer and Minimizing Interference. In Proc. International Conference on Learning Representations

    2019

  16. [24]

    McCloskey, M., & Cohen, N. J. (1989). Catas- trophic Interference in Connectionist Networks: The Sequential Learning Problem. Psychology of Learning and Motivation, 24, 109–165.https: //doi.org/10.1016/S0079-7421(08)60536-8

    1989 doi

  17. [25]

    J., Mirza, M., Xiao, D., Courville, A., & Bengio, Y

    Goodfellow, I. J., Mirza, M., Xiao, D., Courville, A., & Bengio, Y. (2013). An Empirical Investiga- tion of Catastrophic Forgetting in Gradient-Based Neural Networks. arXiv preprint arXiv:1312.6211

    2013 arXiv

  18. [26]

    De Lange, M., Aljundi, R., Masana, M., Parisot, S., Jia, X., Leonardis, A., Slabaugh, G., & Tuyte- laars, T. (2022). A Continual Learning Survey: Defying Forgetting in Classification Tasks. IEEE Trans. Pattern Analysis and Machine Intelli- gence, 44(7), 3366–3385.https://doi.o...

    2022

  19. [27]

    I., Kemker, R., Part, J

    Parisi, G. I., Kemker, R., Part, J. L., Kanan, C., & Wermter, S. (2019). Continual lifelong learn- ing with neural networks: A review. Neural Net- works, 113, 54–71.https://doi.org/10.1016/j. neunet.2019.01.012

    2019 doi

  20. [28]

    K., Ajanthan, T., & Torr, P

    Chaudhry, A., Dokania, P. K., Ajanthan, T., & Torr, P. H. S. (2018). Riemannian Walk for Incremental Learning: Understanding Forgetting and Intransigence. In Proc. European Conference on Computer Vision (pp. 556–572).https://doi. org/10.1007/978-3-030-01252-6_33

    2018 doi

  21. [29]

    Tavallaee, M., Bagheri, E., Lu, W., & Ghorbani, A. A. (2009). A Detailed Analysis of the KDD CUP 99 Data Set. In Proc. IEEE Symp. on Computational Intelligence in Security and De- fense Applications (pp. 1–6).https://doi.org/ 10.1109/CISDA.2009.5356528

    2009

  22. [30]

    H., & Ghorbani, A

    Sharafaldin, I., Lashkari, A. H., & Ghorbani, A. A. (2018). Toward Generating a New Intrusion Detection Dataset and Intrusion Traffic Charac- terization. In Proc. 4th International Conference on Information Systems Security and Privacy (ICISSP) (pp. 108–116)

    2018

  23. [31]

    N., Kaiser, L., & Polosukhin, I

    Vaswani, A., Shazeer, N., Parmar, N., Uszkor- eit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017). Attention Is All You Need. In Advances in Neural Information Processing Systems 30 (pp. 5998–6008)

    2017

  24. [32]

    A., Ibrahim, N

    Shyaa, M. A., Ibrahim, N. F., Zainol, Z., Abdullah, R., Anbar, M., & Alzubaidi, L. (2024). Evolving Cybersecurity Frontiers: A Comprehensive Survey on Concept Drift and Feature Dynamics Aware Machine and Deep Learning in Intrusion Detection Systems. Engineering Applications of...

    2024

  25. [33]

    Chai, Y., Du, L., Qiu, J., Yin, L., & Tian, Z. (2023). Dynamic Prototype Network Based on Sample Adaptation for Few-Shot Malware De- tection. IEEE Transactions on Knowledge and Data Engineering, 35(5), 4754–4766.https:// doi.org/10.1109/TKDE.2022.3142820

    2023

  26. [34]

    Kim, M., Lee, D., Lee, K., Kim, D., Lee, S., & Kim, J. (2022). Deep Sequence Models for Packet Stream Analysis and Early Decisions. In Proc. IEEE 47th Conference on Local Computer Networks (LCN) (pp. 56–63).https://doi.org/ 10.1109/LCN53696.2022.9843272

    2022

  27. [35]

    K., Tadwai, A., Vinta, R., Chan- nappayya, S

    Amalapuram, S. K., Tadwai, A., Vinta, R., Chan- nappayya, S. S., & Tamma, B. R. (2022). Contin- ual learning for anomaly based network intrusion detection. In Proc. 14th International Conference on COMmunication Systems & NETworkS (COM- SNETS) (pp. 497–505). IEEE

    2022

  28. [36]

    Smith,J.S.,Valkov,L.,Halbe,S.,Gutta,V.,Feris, R., Kira, Z., & Karlinsky, L. (2024). Adaptive memory replay for continual learning. In Proc. Ariffin et al.: Preprint Page 22 of 23 Adaptive IDS with Transformers, Continual Learning, and Adversarial Investigation IEEE/CVF Confere...

    2024

  29. [37]

    Guo, J., & Schwaller, P. (2023). Augmented Mem- ory: Capitalizing on Experience Replay to Accel- erate De Novo Molecular Design. ACS Central Science, 9(12), 2273–2285.https://doi.org/10. 1021/acscentsci.3c01480

    2023

  30. [38]

    The MITRE Corporation. (2026). CVE Program Metrics.https://www.cve.org/about/Metrics (accessed 5 July 2026)

    2026

  31. [39]

    Biggio, B., Nelson, B., & Laskov, P. (2012). Poi- soning Attacks against Support Vector Machines. InProc.29thInternationalConferenceonMachine Learning (pp. 1807–1814)

    2012

  32. [40]

    Biggio, B., & Roli, F. (2018). Wild patterns: Ten years after the rise of adversarial machine learning. Pattern Recognition, 84, 317–331.https://doi. org/10.1016/j.patcog.2018.07.023

    2018 doi

  33. [41]

    Umer, M., Dawson, G., & Polikar, R. (2020). Tar- geted Forgetting and False Memory Formation in Continual Learners through Adversarial Backdoor Attacks. In Proc. International Joint Conference on Neural Networks (IJCNN) (pp. 1–8).https: //doi.org/10.1109/IJCNN48605.2020.9206809

    2020

  34. [42]

    Guo, Z., Kumar, A., & Tourani, R. (2024). Per- sistent Backdoor Attacks in Continual Learning. arXiv preprint arXiv:2409.13864

    2024 arXiv

  35. [43]

    Guo, C., Zaki, F., Ariffin, A., & Anuar, N. B. (2025b). Clean-Label Backdoor Attack on CNN-Based Network Traffic Classifiers with RhythmicFlow. In Proc. 2025 IEEE International Conference on Computing (ICOCO) (pp. 447–452).https://doi.org/10.1109/ ICOCO67189.2025.11334106

    2025

  36. [44]

    Apruzzese, G., Andreolini, M., Colajanni, M., & Marchetti, M. (2021). Modeling Realistic Adver- sarialAttacksagainstNetworkIntrusionDetection Systems. Digital Threats: Research and Practice, 3(3), 1–19.https://doi.org/10.1145/3469659

    2021 doi

  37. [45]

    Zhang, X., Wang, Y., Ohtsuki, T., Gui, G., Yuen, C., Di Renzo, M., & Sari, H. (2025). Malware Traffic Classification via Expandable Class Incre- mental Learning With Architecture Search. IEEE Transactions on Information Forensics and Secu- rity, 20, 6074–6085.https://doi.org/1...

    2025

  38. [46]

    Guo, C., Li, X., Cheng, J., Yang, S., & Gong, H. (2025a). Continual Learning for Intrusion Detec- tion Under Evolving Network Threats. Future In- ternet, 17(10), 456.https://doi.org/10.3390/ fi17100456

    2025

  39. [47]

    Chen, H., Wang, Z., Yang, S., Luo, X., He, D., & Chan, S. (2025). Intrusion Detection Using Synaptic Intelligent Convolutional Neural Net- works for Dynamic Internet of Things Envi- ronments. Alexandria Engineering Journal, 111, 78–91.https://doi.org/10.1016/j.aej.2024. 10.014...

    2025 doi

Pith tools

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