Pith. sign in

REVIEW 5 major objections 5 minor 33 references

Neurosymbolic Artificial Intelligence for Robust Network Intrusion Detection: From Scratch to Transfer Learning

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a specific transfer-learning configuration of the ODXU neurosymbolic intrusion-detection framework—reusing a pretrained autoencoder unchanged, retraining the clustering module, and fine-tuning the XGBoost…

desk verdict Useful UQ comparison and a clean ablation design, but the headline transfer-learning claim is untested: every transfer case includes pretrained components and no from-scratch ODXU baseline appears. read the letter →

arxiv 2506.04454 v1 pith:2WXBPKNR submitted 2025-06-04 cs.LG

Huynh T. T. Tran , Jacob Sander , Achraf Cohen , Brian Jalaian , Nathaniel D. Bastian This is my paper
classification cs.LG
keywords neurosymbolicAInetworkintrusiondetectiontransferlearningdeepembeddedclusteringXGBoostuncertaintyquantificationopensetrecognitionmetamodel
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 extends ODXU, a neurosymbolic framework that combines deep embedded clustering, XGBoost symbolic reasoning, and uncertainty quantification, and asks whether pretrained components can be reused across intrusion-detection datasets. On the CIC-IDS-2017 dataset, the from-scratch ODXU model outperforms fully connected and convolutional neural baselines on most metrics, most clearly in false omission rate. The central transfer claim is that reusing the autoencoder 'as is,' training the clustering module from scratch, and fine-tuning the XGBoost classifier (Case 6) gives the best results on ACI-IoT-2023, beating the neural baselines with as few as about 16,000 labeled samples. The paper also claims that metamodel-based uncertainty quantification methods (SHAP, information gain, and probability-augmented) consistently outperform score-based methods (confidence scoring and Shannon entropy) at detecting misclassifications and unknown attack types on both datasets.

What carries the argument

The central object is the ODXU architecture, composed of a deep embedded clustering (DEC) module and an XGBoost classifier. DEC first pretrains a stacked denoising autoencoder on 1500-byte payload vectors extracted by the Payload-Byte tool, then replaces the decoder with a clustering layer optimized by a loss that combines KL divergence, a contrastive term pushing cluster centroids apart, and cross-entropy supervision. XGBoost then operates on the 12-dimensional latent representations to produce interpretable IF-THEN classification rules. The transfer framework defines six cases that vary which of the three components (autoencoder, clustering, classifier) is fine-tuned or trained from scratch, and the uncertainty quantification machinery adds metamodels that take the base classifier's input together with SHAP values, information gain, or sorted class probabilities to estimate prediction confidence.

What would settle it

Train an ODXU model from scratch on ACI-IoT-2023 using the same 16,000-sample training budget and compare its six evaluation metrics—especially multiclass accuracy, false omission rate, and competence—against the Case 6 transfer configuration. If the from-scratch model matches or exceeds the transfer model, the paper's central transfer-learning conclusion is unsupported.

Watch

Extended reading notes

Core claim

ODXU—an architecture in which a deep embedded clustering network learns a low-dimensional latent representation of raw packet payloads and XGBoost classifies those latent features—can be transferred across intrusion-detection datasets by selectively reusing its components. Through an ablation over six transfer configurations, the paper finds that the best recipe on ACI-IoT-2023 is to load a pretrained autoencoder and leave it frozen, reinitialize and train the clustering layer on the target data, and fine-tune the pretrained XGBoost classifier (Case 6). With this configuration, the model achieves the best score on all six evaluation metrics—multiclass accuracy .983, binary accuracy .987, misclassified positive rate .019, false omission rate .014, F1 .988, and competence .969—against FcNN and 1D-CNN baselines, and it reaches that performance with roughly 50% of the target training data. The paper further reports that metamodel-based uncertainty quantification, which trains a secondary classifier to predict whether the base model is correct, outperforms simple score-based methods for both misclassification detection and open set recognition.

Load-bearing premise

The paper assumes that the observed performance of the transfer cases reflects a benefit of knowledge transferred from CIC-IDS-2017, because every transfer case includes at least one pretrained component and no ODXU model trained entirely from scratch on ACI-IoT-2023 is evaluated as a baseline.

Editorial extensions

If this is right

  • If the transfer recipe holds, deploying an intrusion-detection system on a new network environment could require only a fraction of the labeled data that a from-scratch model needs, since the autoencoder's learned payload representations are reusable across datasets.
  • The Case 6 configuration—frozen autoencoder, retrained clustering, fine-tuned classifier—provides a concrete starting point for adapting ODXU to other network traffic datasets with limited labels.
  • Metamodel-based uncertainty estimates, especially SHAP-based ones for open set recognition, could let operators flag novel attacks with a single score threshold rather than relying on raw predicted probabilities.
  • Early stopping with conservative thresholds (20 rounds, loss deltas of 0.0005 and 0.005) preserves accuracy while cutting training time, making the approach more practical for operational deployment.
  • The consistent advantage of fine-tuning the XGBoost classifier across all cases suggests that classifier initialization matters more than autoencoder fine-tuning when transferring between these two datasets.

Reading between the lines

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

  • The paper's transfer conclusion would be stronger if it included a zero-transfer baseline: an ODXU model trained entirely from scratch on ACI-IoT-2023 with the same data budget. Without that baseline, the observed gains could stem from the architecture itself or from the target-data training rather than from knowledge transferred from CIC-IDS-2017.
  • The frozen-autoencoder result hints that the 12-dimensional latent representation of packet payloads captures transferable structure across quite different network environments (enterprise versus IoT), which could be tested directly by probing whether latent neighborhoods align between the two datasets.
  • Because the metamodel superiority holds on both datasets and for both tasks, a practical recommendation—not explicitly stated in the paper—would be to prefer metamodel-based uncertainty scores whenever a labeled validation set is available for training the metamodel.
  • The 50% data threshold (about 16,000 samples) is reported for accuracy only; an extension would be to measure how the transfer configuration's advantage in false omission rate and competence changes as the data budget shrinks below 50%.
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

5 major / 5 minor

Summary. The paper extends the ODXU neurosymbolic framework for network intrusion detection, integrating deep embedded clustering, XGBoost, and several uncertainty quantification (UQ) methods. It evaluates ODXU on CIC-IDS-2017 and proposes a transfer-learning strategy with six configurations over the autoencoder, clustering, and classifier components, evaluated on ACI-IoT-2023 under different training-data portions. The headline claims are that the best transfer configuration (Case 6, Experiment 6) outperforms FcNN and 1D-CNN baselines on ACI-IoT-2023, and that metamodel-based UQ methods (SHAP, Information Gain, probability-based) consistently outperform score-based methods (confidence, entropy) for misclassification detection and open-set recognition on both datasets.

Significance. If the transfer-learning and UQ claims are established, the work would be a useful contribution to neurosymbolic NIDS, particularly for data-efficient adaptation to new traffic domains and for uncertainty-aware operation. The paper has concrete strengths: it evaluates on an external dataset (ACI-IoT-2023), provides an ablation over transfer configurations, reports multiple evaluation metrics, and includes open-set-recognition experiments. However, the central transfer-learning benefit is not isolated by the current experimental design, and the UQ 'consistent superiority' claim is partly contradicted by the reported numbers. The contribution is therefore potentially valuable but requires additional experiments and analysis.

major comments (5)
  1. [§4.3, Table 1; §5.2, Table 5] The transfer-learning claim is not directly tested because every one of the six cases in Table 1 uses the CIC-IDS-2017 pretrained autoencoder (either as-is or fine-tuned), and the clustering module is always initialized from the pretrained encoder parameters per the footnote. No ODXU model trained entirely from scratch on ACI-IoT-2023 is evaluated; a model with random AE initialization, clustering trained from scratch, and classifier trained from scratch is absent. Consequently, the reported superiority of Case 6 over FcNN/1D-CNN could be due to the ODXU architecture or the Payload-Byte preprocessing pipeline rather than to cross-dataset pretraining. Adding a zero-transfer baseline with identical data portions and hyperparameters is necessary to support the abstract's and conclusion's transfer-learning claim.
  2. [§5.2, Table 5] All accuracy values in Table 5 appear to be single runs, with no standard deviations, confidence intervals, or significance tests. The ranking of configurations relies on differences as small as .0014 (e.g., Case 6 vs. Case 3 at 50% training data), which may be within run-to-run variation. The paper should report results over multiple seeds with variance or, at minimum, justify that the observed differences exceed typical training noise.
  3. [§4.3 and §5.2] The FcNN and 1D-CNN baselines on ACI-IoT-2023 are compared against transfer cases at 10-75% training data, but the paper does not state what portion of the target data these baselines were trained on; the fixed baseline values .9808 and .9679 in Section 5.2 are ambiguous. In addition, the ACI test set is heavily rebalanced (benign downsampled by 95%, rare classes upsampled by 200%), so the reported accuracies do not reflect the original class distribution. The operational relevance of the 'outperforms' claim should be demonstrated on the original distribution or explicitly discussed as rebalanced-distribution accuracy.
  4. [§5.3.2, Table 9] The general claim that 'metamodel-based approaches consistently outperform score-based ones' (Section 5.3 and Conclusion) is contradicted by the CIC-IDS-2017 OSR results: at TP@(TN=.95), Confidence and Entropy reach .212, while MetaUQProb, MetaUQSHAP, and MetaUQIG reach only .145, .145, and .142. The paper should qualify the claim to specific metrics or datasets, or provide an explanation for this reversal before asserting consistent superiority.
  5. [§5.3, Tables 8-9] The UQ comparisons are also based on single-run AUROC and TP@(TN=.95) values, with no uncertainty estimates or significance tests. Several differences are small (e.g., .923 vs. .925 on CIC-IDS-2017 misclassification detection), so the claim of consistent superiority of metamodel-based methods needs error bars or statistical testing to be convincing.
minor comments (5)
  1. [§5.3.1] The sentence 'MetaUQProb, MetaUQIG and MetaUQSHAP achieve AUROCs of .923 and .925, respectively' lists three methods but only two values; the third value should be stated explicitly (see Table 8).
  2. [§4.4] The text says the metamodel training set is formed by subsampling class 0 to five times the size of class 1 and then calls this a 'balanced' split; a 5:1 ratio is imbalanced, so the wording should be corrected.
  3. [Figure 3 caption] Averaging the 'FT' and 'Train' clustering accuracies for the 'As is' autoencoder settings obscures the underlying data; plotting all points or using a standard interaction plot with raw values would be clearer.
  4. [Eq. (2b)] The contrastive loss notation sums ∥u_i − u_j∥² over unspecified index pairs; if i=j is included, the term is zero and the expression is undefined when all centroids coincide. The summation index should be defined explicitly.
  5. [§4.3] The phrase 'This reduced set was further split equally into XGBoost-Train and XGBoost-Test' is ambiguous: it could refer to the reduced DEC-Test or to the full reduced dataset. Clarify which set is split.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the UQ metamodeling is a standard supervised setup and the transfer-learning claim, though underdetermined by a missing from-scratch control, is not forced by construction.

full rationale

The paper's load-bearing empirical claims are evaluated on external benchmarks (CIC-IDS-2017 and ACI-IoT-2023) against independent FcNN and 1D-CNN baselines, so the reported gains are not equivalent to the inputs by construction. The UQ metamodel is a secondary binary classifier trained on correctness labels defined in Eq. (3); its misclassification AUROC is computed on a held-out 20% of the metamodel data, which is a standard supervised evaluation rather than a circular reduction. Score-based UQ methods are fixed heuristics, and the metamodel's advantage is an empirical, potentially unfair but non-circular, comparison. Self-citations to ODXU [2] and Payload-Byte [10] supply the framework and preprocessing tooling, but both are implemented and tested here against external data, so the citations are not load-bearing in a circular sense. The main validity concern is that every transfer case in Table 1 uses at least one pretrained component, and no ODXU-from-scratch baseline on ACI-IoT-2023 is reported, so the specific contribution of cross-dataset pretraining is underdetermined; however, this is a missing-control/confounding issue, not a case where a prediction reduces to a fit or to a self-citation by definition.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a large set of hyperparameter choices (latent dimension, downsampling ratios, early stopping thresholds, metamodel subsampling) and domain assumptions, but no new physical or conceptual entities are introduced. The most consequential assumptions are the comparability of Payload-Byte vectors across datasets and the validity of uncertainty scores for unknown attack detection.

free parameters (6)
  • Latent dimension = 12
    DEC compresses 1500-byte payloads to a 12-dimensional latent space (Section 4.3). No ablation is shown for this choice.
  • Downsampling ratios (CIC-IDS-2017) = Benign and DoS Hulk 90%, DDoS 67%
    Section 4.3 rebalances the training/tuning set; this alters the test distribution and affects all reported metrics.
  • Downsampling/upsampling ratios (ACI-IoT-2023) = Benign 95% downsample; ICMP and UDP Flood 200% upsample
    Section 4.3 rebalances the dataset before the 50/50 split, changing class priors for both training and evaluation.
  • Early stopping hyperparameters = eta=20, delta_AE=0.0005, delta_Cluster=0.005
    Section 5.2 selects these by grid search (Table 6) as the best configuration; they are fitted to the validation set.
  • Metamodel subsampling ratio = 5:1 correct-to-incorrect
    Section 4.4 subsamples class 0 to five times class 1 when training the UQ metamodel; this ratio is chosen without ablation.
  • Number of centroids nc = Not reported
    The contrastive loss (Eq. 2b) depends on nc, but the paper never states how many cluster centroids are used in DEC.
assumptions (5)
  • domain assumption Payload-Byte conversion yields a standardized 1500-byte feature vector independent of protocol and comparable across datasets.
    Section 3.1 states Payload-Byte converts each packet payload into a standardized 1500-byte vector; transfer between CIC-IDS-2017 and ACI-IoT-2023 relies on this shared input representation.
  • domain assumption XGBoost decision paths can be treated as propositional logic expressions for symbolic reasoning.
    Section 2.2 asserts tree-based models offer a natural integration point for propositional logic; this frames XGBoost as the symbolic component of ODXU.
  • domain assumption Unknown attack types can be flagged via uncertainty scores without retraining the base classifier.
    Section 5.3.2 evaluates OSR by applying uncertainty thresholds; this assumes the uncertainty signal trained on known classes generalizes to held-out attack types.
  • ad hoc to paper The DEC objective with added contrastive and cross-entropy losses improves clustering for intrusion detection.
    Section 3.2 introduces the three-term loss (Eq. 1) as a modification to standard DEC; no comparison against plain DEC is provided, so the benefit of this design is assumed.
  • standard math Standard ML background: KL divergence, cross-entropy, gradient boosting, and SHAP formulas are assumed correct.
    Equations (2a)-(2c), (7), and (9) rely on standard definitions from prior literature.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Neurosymbolic Artificial Intelligence for Robust Network Intrusion Detection: From Scratch to Transfer Learning." pith.science (2026). https://pith.science/paper/2WXBPKNR

@misc{pith2026250604454,
  author       = {Pith},
  title        = {Pith review of: Neurosymbolic Artificial Intelligence for Robust Network Intrusion Detection: From Scratch to Transfer Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2WXBPKNR}},
  note         = {Machine review of arXiv:2506.04454}
}
read the original abstract

Network Intrusion Detection Systems (NIDS) play a vital role in protecting digital infrastructures against increasingly sophisticated cyber threats. In this paper, we extend ODXU, a Neurosymbolic AI (NSAI) framework that integrates deep embedded clustering for feature extraction, symbolic reasoning using XGBoost, and comprehensive uncertainty quantification (UQ) to enhance robustness, interpretability, and generalization in NIDS. The extended ODXU incorporates score-based methods (e.g., Confidence Scoring, Shannon Entropy) and metamodel-based techniques, including SHAP values and Information Gain, to assess the reliability of predictions. Experimental results on the CIC-IDS-2017 dataset show that ODXU outperforms traditional neural models across six evaluation metrics, including classification accuracy and false omission rate. While transfer learning has seen widespread adoption in fields such as computer vision and natural language processing, its potential in cybersecurity has not been thoroughly explored. To bridge this gap, we develop a transfer learning strategy that enables the reuse of a pre-trained ODXU model on a different dataset. Our ablation study on ACI-IoT-2023 demonstrates that the optimal transfer configuration involves reusing the pre-trained autoencoder, retraining the clustering module, and fine-tuning the XGBoost classifier, and outperforms traditional neural models when trained with as few as 16,000 samples (approximately 50% of the training data). Additionally, results show that metamodel-based UQ methods consistently outperform score-based approaches on both datasets.

Figures

Figures reproduced from arXiv: 2506.04454 by the authors.

Figure 1
Figure 1. The architecture of the ODXU model Input Data Encoder x Decoder Reconstructed Input Phase I: Autoencoder Pretraining Input Data Encoder x Clustering Latent Representations Phase II: Clustering Optimization (a) DEC model structure Samples (x) Metamodel (m) Certainty score (z) Base Classifier (b) Class probabilities (p) (b) Metamodel inference [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. General structure of (a) DEC and (b) metamodel [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The Accuracy Interaction Plot for different configurations of the ODXU model. For “FT” AutoEncoder [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Uncertainty score distributions in the MetaUQ [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Comparison of uncertainty score distributions. Column 1: CIC-IDS-2017. Column 2: ACI-IoT-2023. Row 1: [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 26 canonical work pages

  1. [1]

    An intelligent tree-based intrusion detection model for cyber security.Journal of Network and Systems Management, 29(2):20, 2021

    Mohammad Al-Omari, Majdi Rawashdeh, Fadi Qutaishat, Mohammad Alshira’H, and Nedal Ababneh. An intelligent tree-based intrusion detection model for cyber security.Journal of Network and Systems Management, 29(2):20, 2021

    work page 2021

  2. [2]

    Jacob Sander, Chung-En Johnny Yu, Brian Jalaian, and Nathaniel D. Bastian. Uncertainty-quantified neu- rosymbolic ai for open set recognition in network intrusion detection. In2024 IEEE Military Communications Conference (MILCOM), pages 13–18, Washington, DC, USA, 2024. IEEE

    work page 2024

  3. [3]

    Brian Jalaian and Nathaniel D. Bastian. Neurosymbolic ai in cybersecurity: Bridging pattern recognition and symbolic reasoning. InProceedings of the 2023 IEEE Military Communications Conference (MILCOM), pages 268–273, Boston, MA, USA, 2023. IEEE

    work page 2023

  4. [4]

    Application of image processing and transfer learning for the detection of rust disease.Scientific Reports, 13(1):5133, 2023

    Fereshteh Shahoveisi, Hamed Taheri Gorji, Seyedmojtaba Shahabi, Seyedali Hosseinirad, Samuel Markell, and Fartash Vasefi. Application of image processing and transfer learning for the detection of rust disease.Scientific Reports, 13(1):5133, 2023

    work page 2023

  5. [5]

    Shahin Amiriparian, Tobias Hübner, Vincent Karas, Maurice Gerczuk, Sandra Ottl, and Björn W Schuller. Deepspectrumlite: A power-efficient transfer learning framework for embedded speech and audio processing from decentralized data.Frontiers in Artificial Intelligence, 5:856232, 2022

    work page 2022

  6. [6]

    Multimodal transfer deep learning with applications in audio-visual recognition.https://arxiv.org/abs/1412.3121, 2014

    Seungwhan Moon, Suyoun Kim, and Haohan Wang. Multimodal transfer deep learning with applications in audio-visual recognition.https://arxiv.org/abs/1412.3121, 2014. arXiv:1412.3121 [cs.LG]

    arXiv 2014

  7. [7]

    Transfer learning for medical image classification: a literature review.BMC medical imaging, 22(1):69, 2022

    Hee E Kim, Alejandro Cosa-Linan, Nandhini Santhanam, Mahboubeh Jannesari, Mate E Maros, and Thomas Ganslandt. Transfer learning for medical image classification: a literature review.BMC medical imaging, 22(1):69, 2022

    work page 2022

  8. [8]

    An intrusion-detection model.IEEE Transactions on software engineering, SE-13(2):222– 232, 1987

    Dorothy E Denning. An intrusion-detection model.IEEE Transactions on software engineering, SE-13(2):222– 232, 1987

    work page 1987

Show all 33 references
  1. [9]

    Generalized out-of-distribution detection: A survey

    Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of-distribution detection: A survey. International Journal of Computer Vision, 132(12):5635–5662, Jun 2024

    2024

  2. [10]

    Pavlik, and Nathaniel D

    Yasir Ali Farrukh, Irfan Khan, Syed Wali, David Bierbrauer, John A. Pavlik, and Nathaniel D. Bastian. Payload- byte: A tool for extracting and labeling packet capture files of modern network intrusion detection datasets. InProceedings of the 2022 IEEE/ACM International Confere...

    2022

  3. [11]

    Unsw-nb15: A comprehensive data set for network intrusion detection systems (unsw-nb15 network data set)

    Nour Moustafa and Jill Slay. Unsw-nb15: A comprehensive data set for network intrusion detection systems (unsw-nb15 network data set). InProceedings of the 2015 Military Communications and Information Systems Conference (MilCIS), pages 1–6, Canberra, ACT, Australia, 2015. IEEE

    2015

  4. [12]

    Toward generating a new intrusion detection dataset and intrusion traffic characterization.ICISSp, 1:108–116, 2018

    Iman Sharafaldin, Arash Habibi Lashkari, Ali A Ghorbani, et al. Toward generating a new intrusion detection dataset and intrusion traffic characterization.ICISSp, 1:108–116, 2018

    2018

  5. [13]

    Aci iot network traffic dataset 2023, Apr 2024

    Emily Nack. Aci iot network traffic dataset 2023, Apr 2024

    2023

  6. [14]

    Alice Bizzarri, Chung-En Yu, Brian Jalaian, Fabrizio Riguzzi, and Nathaniel D. Bastian. A synergistic ap- proach in network intrusion detection by neurosymbolic ai. https://arxiv.org/abs/2406.00938, 2024. arXiv:2406.00938 [cs.CR]

    2024 arXiv

  7. [15]

    Chen, Fahim Tajwar, Ananya Kumar, Huaxiu Yao, Percy Liang, and Chelsea Finn

    Yoonho Lee, Annie S. Chen, Fahim Tajwar, Ananya Kumar, Huaxiu Yao, Percy Liang, and Chelsea Finn. Surgical fine-tuning improves adaptation to distribution shifts. https://arxiv.org/abs/2210.11466, 2022. arXiv:2210.11466 [cs.LG]

    2022 arXiv

  8. [16]

    Transtailor: Pruning the pre-trained model for improved transfer learning

    Bingyan Liu, Yifeng Cai, Yao Guo, and Xiangqun Chen. Transtailor: Pruning the pre-trained model for improved transfer learning. InProceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 8627–8634, Vancouver, BC, Canada (held virtually), 2021. AAAI Press

    2021

  9. [17]

    Spottune: Transfer learning through adaptive fine-tuning

    Yunhui Guo, Honghui Shi, Abhishek Kumar, Kristen Grauman, Tajana Rosing, and Rogerio Feris. Spottune: Transfer learning through adaptive fine-tuning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4805–4814, Long Beach, CA, US...

    2019

  10. [19]

    A baseline for detecting misclassified and out-of-distribution examples in neural networks.https://arxiv.org/abs/1610.02136, 2016

    Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks.https://arxiv.org/abs/1610.02136, 2016. arXiv:1610.02136 [cs.LG]

    2016 arXiv

  11. [20]

    Wong, Alexander M

    Joshua A. Wong, Alexander M. Berenbeim, David A. Bierbrauer, and Nathaniel D. Bastian. Uncertainty-quantified, robust deep learning for network intrusion detection. InProceedings of the 2023 Winter Simulation Conference (WSC), pages 2470–2481, San Antonio, TX, USA, 2023. IEEE

    2023

  12. [21]

    Vera Liao, Karthikeyan Natesan Ramamurthy, Jiri Navratil, Prasanna Sattigeri, Kush R

    Soumya Ghosh, Q. Vera Liao, Karthikeyan Natesan Ramamurthy, Jiri Navratil, Prasanna Sattigeri, Kush R. Varsh- ney, and Yunfeng Zhang. Uncertainty quantification 360: A holistic toolkit for quantifying and communicating the uncertainty of ai.https://arxiv.org/abs/2106.01410, 20...

    2021 arXiv

  13. [22]

    Confidence scoring using whitebox meta-models with linear classifier probes

    Tongfei Chen, Jiˇrí Navrátil, Vijay Iyengar, and Karthikeyan Shanmugam. Confidence scoring using whitebox meta-models with linear classifier probes. InProceedings of the 22nd International Conference on Artificial Intelligence and Statistics (AISTATS 2019), pages 1467–1475, Na...

    2019

  14. [23]

    Lundberg and Su-In Lee

    Scott M. Lundberg and Su-In Lee. A unified approach to interpreting model predictions.Advances in Neural Information Processing Systems, 30:4765–4774, 2017

    2017

  15. [24]

    Ross Quinlan

    J. Ross Quinlan. Induction of decision trees.Machine learning, 1:81–106, 1986

    1986

  16. [25]

    Unsupervised deep embedding for clustering analysis

    Junyuan Xie, Ross Girshick, and Ali Farhadi. Unsupervised deep embedding for clustering analysis. InProceedings of the 33rd International Conference on Machine Learning (ICML), ICML’16, pages 478–487, New York, NY , USA, 2016. JMLR.org

    2016

  17. [26]

    Xgboost: A scalable tree boosting system

    Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. InProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD ’16), pages 785–794, San Francisco, CA, USA, 2016. ACM

    2016

  18. [27]

    Quantifying information flow using min-entropy

    Geoffrey Smith. Quantifying information flow using min-entropy. InProceedings of the 2011 Eighth International Conference on Quantitative Evaluation of Systems (QEST), pages 159–167, Aachen, Germany, 2011. IEEE

    2011

  19. [28]

    Single-model uncertainties for deep learning.Advances in Neural Information Processing Systems, 32:6415–6425, 2019

    Natasa Tagasovska and David Lopez-Paz. Single-model uncertainties for deep learning.Advances in Neural Information Processing Systems, 32:6415–6425, 2019

    2019

  20. [29]

    Lundberg, Gabriel G

    Scott M. Lundberg, Gabriel G. Erion, and Su-In Lee. Consistent individualized feature attribution for tree ensembles, 2019

    2019

  21. [30]

    From local explanations to global understanding with explainable ai for trees.Nature machine intelligence, 2(1):56–67, 2020

    Scott M Lundberg, Gabriel Erion, Hugh Chen, Alex DeGrave, Jordan M Prutkin, Bala Nair, Ronit Katz, Jonathan Himmelfarb, Nisha Bansal, and Su-In Lee. From local explanations to global understanding with explainable ai for trees.Nature machine intelligence, 2(1):56–67, 2020

    2020

  22. [31]

    The use of the area under the roc curve in the evaluation of machine learning algorithms

    Andrew P Bradley. The use of the area under the roc curve in the evaluation of machine learning algorithms. Pattern recognition, 30(7):1145–1159, 1997. Neurosymbolic Artificial Intelligence for Robust Network Intrusion Detection: From Scratch to Transfer Learning

    1997

  23. [32]

    Bastian, Daniel Clouse, Bradford Kline, and Susmit Jha

    Brian Matejek, Ashish Gehani, Nathaniel D. Bastian, Daniel Clouse, Bradford Kline, and Susmit Jha. Safeguarding network intrusion detection models from zero-day attacks and concept drift. InProceedings of the AAAI Workshop on Artificial Intelligence for Cyber Security (AICS), ...

    2024

  24. [33]

    ACI IoT Network Traffic Dataset 2023

    Nathaniel Bastian, David Bierbrauer, Morgan McKenzie, and Emily Nack. ACI IoT Network Traffic Dataset 2023. https://dx.doi.org/10.21227/qacj-3x32, 2023. DOI: 10.21227/qacj-3x32. A Per-class accuracy Table 10: Per-class accuracy across models of the CIC-IDS-2017. Class Accuracy...

    2023 doi

  25. [2020]

    arXiv:2007.04466 [cs.LG]

    2007 arXiv

Pith tools

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