Pith. sign in

REVIEW 4 major objections 7 minor 24 references

A passive observer can classify quantum identity-authentication stages from photon timing and optical power alone, at up to 98% accuracy, without collapsing the quantum state.

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 · grok-4.5

2026-07-31 09:26 UTC pith:PT5OYQOI

load-bearing objection Solid lab demo that QIA data/auth stages leak via passive optical timing and power at 10–30% taps, but the signal is probably their H/V-only vs HWP-switched implementation, not an inherent protocol property. the 4 major comments →

arxiv 2607.24639 v1 pith:PT5OYQOI submitted 2026-07-27 quant-ph cs.NI

Experimental Side Channel Analysis of Protocol Stages in Quantum Identity Authentication

Marwan Elawady , Lance Young , Contessa Wilburn , Blaine Keyton , Carrie Houston , Mohamed Shaban , Muhammad Ismail This is my paper
classification quant-ph cs.NI
keywords protocol stage identificationquantum identity authenticationside channel analysispassive eavesdroppingphysical layer securityquantum network vulnerabilityphoton timingoptical power
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.

Quantum identity authentication hides authentication qubits among data qubits using a shared secret so that man-in-the-middle attacks can be detected. This paper shows that those stages still leave distinct physical fingerprints. On an entangled-photon testbed, a beam splitter diverted 10% or 30% of the light to a passive observer who recorded only photon arrival times and optical power; no quantum-state measurement was performed. From engineered timing and intensity features, machine-learning models separated data-transmission intervals from authentication intervals at 96–98% accuracy while Bell correlations remained intact. An adversary who knows the stage can therefore skip authentication qubits and harvest data qubits, defeating the protocol’s security without ever touching the quantum information itself. The work argues that QIA designs must treat physical-layer stage leakage as a first-class threat.

Core claim

On a polarization-entangled photon link, passive side-channel observations of photon arrival timing and optical power are sufficient to classify quantum identity-authentication protocol stages (data transmission versus authentication) at up to 98% accuracy (F1 97%) with a 30% tap and 96% accuracy (F1 94%) with a 10% tap, while the CHSH value stays above 2 so the quantum channel is not collapsed.

What carries the argument

Engineered side-channel features—inter-arrival-time statistics, count-rate interactions, irradiance variability, and cross-modal power–timing products—fed to supervised classifiers (Random Forest, LSTM, GRU) that map passive observations X to stage label S.

Load-bearing premise

The outsider is assumed to already possess enough labeled traces of known data versus authentication intervals with which to train the classifiers—labels that the shared secret would normally keep hidden in the field.

What would settle it

Give the observer only continuous unlabeled traces with no experimenter-supplied stage boundaries or key-derived schedules; if classification then falls to chance, the claimed field attack does not hold.

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

If this is right

  • Malicious intermediate nodes can extract data qubits while leaving authentication qubits untouched, enabling undetectable man-in-the-middle attacks.
  • Even a 10% optical tap already yields >90% stage accuracy, so low-impact physical access is sufficient.
  • Timing features dominate and power features are complementary, so defenses must mask both photon statistics and intensity.
  • QIA and similar quantum-network protocols need stage indistinguishability at the physical layer, not only at the protocol layer.
  • Brighter operating points improve leakage visibility, so practical high-rate links may be more exposed rather than less.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same timing-and-power feature pipeline could fingerprint other multi-stage quantum protocols whose hardware duty cycles differ, such as QKD sifting versus key generation.
  • Constant-rate dummy emission or randomized attenuation during authentication are natural countermeasures the paper’s results immediately suggest testing.
  • If unsupervised or few-shot methods recover stage structure from unlabeled traces, the labeling assumption weakens and the practical threat grows.
  • Classical side-channel hygiene—power equalization and timing padding—may transfer directly to quantum optical links.

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. The paper reports an experimental study on a polarization-entangled SPDC testbed of whether a passive optical tap (10% or 30% beam-splitter sampling) leaking photon-arrival timing (time tagger/SPAD) and optical power (power meter) allows a machine-learning classifier to distinguish the data-transmission and authentication stages of the interleaved QIA protocol of [19]. The authors verify that the tap preserves CHSH violation (S > 2 in all configurations), quantify tap-induced reductions in singles and coincidence rates, engineer timing/intensity/interaction features, and train Random Forest, LSTM, and GRU classifiers on experimenter-labeled segments from key-permuted sessions at two pump currents. The best configuration (Random Forest, combined features, 43 mA) reaches 98.1% accuracy / 97.0% F1 at 30% tapping and 96.3% / 94.2% at 10% tapping. The authors conclude that QIA stage inference is feasible under constrained passive access and that this undermines QIA's protection against man-in-the-middle attacks.

Significance. If the result holds, this is a useful early demonstration that protocol-level structure in quantum communication can leak through passive optical observation — a threat class largely unaddressed in the QIA/QKD physical-layer security literature, which has concentrated on bit-level and electronics-emission leakage. Strengths worth naming: the work is experimental rather than simulation-based, on a polarization-entangled SPDC testbed with a CHSH witness (S = 2.32–2.36 across tap ratios) verifying non-disturbance; it quantifies the observability/stealth tradeoff via measured photon-count and coincidence-rate reductions at 10% and 30% tapping (Fig. 2); it compares three model families and three feature modalities at two pump currents (Tables I–II); and it includes confusion matrices and permutation-importance analysis, giving some interpretability. The classification itself is ordinary supervised learning with experimenter-assigned labels, so there is no circularity concern of the kind that afflicts parameter-fitted derivations. The significance is, however, bounded by two open questions: whether the separability is inherent to QIA or an artifact of this implementation's H/V-only data,

major comments (4)
  1. [§V.B, Table II; §IV.A] The claim that the leakage is 'driven by inherent protocol behavior rather than specific operating conditions' is not adequately supported, and the paper's own description points to a more mundane mechanism. A polarization-insensitive beam-splitter tap plus SPAD/power meter measures only arrival statistics and intensity; the marginal detection process at the tap is identical for H, V, D, and A photons. Therefore the discriminative signal cannot be the quantum states themselves and must arise from preparation-side dynamics that differ between stages. Sec. IV.A builds in exactly such an asymmetry: data transmission uses only H and V chosen randomly, while authentication uses four states with the X basis realized by physically repositioning the HWP to 22.5 deg. Stage transitions therefore coincide with mechanical switching, altered state-switching sequences, and possibly different D/A vs H/
  2. [§IV.E; §IV.B; §IV.D] The evaluation protocol risks temporal leakage that could inflate the headline accuracies (98%/96%). The pipeline uses rolling-window statistics and lag/delta features (Sec. IV.D) computed over traces that are then regrouped into labeled segments (Sec. IV.B), after which an 80/20 split is applied (Sec. IV.E). If the split is at the sample level rather than the session/run level, adjacent or overlapping windows from the same session appear in both train and test, and autocorrelation plus rolling-window overlap will leak stage information. Similarly, SMOTE and the F1-based classification-threshold optimization must be confined to the training/validation folds; the text does not state where SMOTE is applied or on which set the threshold is tuned. The manuscript also does not report how many distinct sessions or experimental runs underlie the 9,000 samples per configuration, so cross-session
  3. [§III.B; §IV.B; §IV.E] The threat model states the observer 'has no prior knowledge of the transmitted data, shared secret keys, or protocol parameters,' yet the classifiers are supervised and are trained on segments labeled by the experimenters using knowledge of the protocol schedule and key permutations (Sec. IV.B). The paper never explains how a real passive adversary would obtain comparably labeled training traces in the field — without the shared key, the adversary does not know which intervals are authentication rounds, which is precisely the information the protocol is designed to hide. This does not invalidate the demonstration of separability, but it is load-bearing for the claim that this constitutes a practical attack. The authors should address label acquisition directly: e.g., show that unsupervised clustering or a small amount of known-schedule traffic suffices to bootstrap labels, or demonstrat
  4. [§III.B; §V.D; Fig. 2] The security conclusion — that stage inference 'renders the QIA ineffective and makes undetectable MitM attacks feasible' — is not supported by an end-to-end attack analysis. A passive tap at 10–30% diverts only a fraction of each photon; learning stage labels from the tap does not by itself let the adversary read the data qubits, since extracting the data stream requires intercepting (and consuming) the data photons, which induces loss that Bob can in principle detect, and the observed count-rate reductions (29% at 30% tap, ~6% at 10% tap, Fig. 2) are themselves a potential tell. Moreover, the data stage transmits only H/V states (Sec. IV.A), so the 'data' here is effectively classical bits in a single basis; the implications for a QIA deployment where data qubits use both bases are not discussed. The discussion should sketch the complete attack chain (tap + selective interception + for
minor comments (7)
  1. [§IV.E; §III; Fig. 1] Typo: 'while labels 2 = 1 represents the authentication stage' should read 'label s2 = 1'. Also 'used t facilitate' (Sec. III opening) should be 'used to facilitate', and 'coincidence-based' is misspelled as 'coincidence-based' ('coincidence' typo) in the Fig. 1 caption.
  2. [§V; §IV.B] The number of experimental sessions/runs and the duration of each trace underlying the 9,000 samples per configuration should be reported, along with the window length used for rolling/lag features; these are needed to judge sample independence and reproducibility.
  3. [Fig. 2; §V.A] Fig. 2: at 10% tapping the coincidence rate drops only 2.23% while the singles count drops 5.97%; a brief physical explanation of this asymmetry (e.g., heralding geometry, which arm is tapped) would help readers interpret the detectability tradeoff discussed in Sec. V.A.
  4. [§IV.D.2, Eq. (5)] Eq. (5): the regularization constant 0.01 in the IAT-regularity definition is unmotivated; please state its role and whether results are sensitive to it, particularly since IAT-regularity features rank among the dominant features in Figs. 5–6.
  5. [§V.A] The CHSH values (2.324–2.359) are reported with uncertainties but the number of measurement settings, integration time, and whether the tap was present during the baseline (no-sampling) measurement should be specified; also clarify whether 'no sampling' means no beam splitter or a 0% tap, since inserting and removing optics can change alignment.
  6. [§IV.E; Tables I–II] The LSTM/GRU input structure is unclear: are sequences formed from the engineered feature vectors, and if so, what is the sequence length? Given that Random Forest on static features outperforms the sequence models throughout Tables I–II, a sentence on why the recurrent models underperform (e.g., sequences too short, features already encoding temporal structure) would strengthen Sec. V.B.
  7. [§IV] Code and data availability is not stated. Given that the contribution is an empirical ML pipeline, releasing the feature-extraction code and anonymized traces (or at least the train/test split scripts) would substantially improve verifiability.

Circularity Check

0 steps flagged

Empirical supervised classification study; no derivation that folds the target into its inputs

full rationale

The paper’s central claim is experimental: passive optical-power and photon-timing observations, collected via a beam splitter on a polarization-entangled testbed, suffice for ML classifiers to distinguish QIA data-transmission vs authentication stages at high accuracy (up to 98%/F1 97% at 30% tap; 96%/F1 94% at 10% tap), while CHSH remains >2. Labels are assigned from the known experimental schedule (key-permuted sessions with known interleaving of data and auth rounds); features are engineered from raw SPAD timestamps and power-meter readings; models (RF, LSTM, GRU) are trained and evaluated on held-out splits with SMOTE and stratified CV. That is ordinary supervised learning, not a first-principles derivation whose output is forced by construction from a fitted definition. The only self-reference is implementation of the authors’ prior QIA design [19]; that citation supplies the protocol under test, not a uniqueness theorem or ansatz that makes the side-channel accuracies tautological. No equation equates a predicted quantity to a fitted input; no load-bearing uniqueness is imported; no known empirical law is merely renamed. Correctness concerns (e.g., whether separability is an HWP/preparation artifact rather than inherent stage leakage) are outside the circularity criterion. Score 0; steps empty.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claim rests on standard quantum-optics and ML practice plus a few domain choices: the specific interleaved QIA protocol of [19], the assumption that a beam-splitter tap yields usable classical side channels without destroying entanglement (empirically checked via CHSH), and the supervised-learning setup that treats experimenter stage labels as ground truth. No new physical entities are postulated. Free parameters are experimental knobs and ML hyperparameters, not constants fitted to force a theory curve.

free parameters (5)
  • Optical tapping ratio (10% / 30%) = 10% and 30%
    Chosen experimental access levels that directly set how much side-channel signal the observer receives and how much the main channel is attenuated.
  • Pump laser current (38 mA / 43 mA) = 38 mA and 43 mA
    Operating points chosen to vary brightness; classification tables are reported separately at each setting.
  • Random Forest hyperparameters (n_trees=500, min_leaf=2) = 500 trees, min leaf 2
    Hand-set model capacity used for the best-reported accuracies.
  • LSTM/GRU learning rate, dropout, focal-loss gamma, class threshold = lr=1e-4, dropout=0.4, γ=2
    Tuned on validation data (lr=1e-4, dropout=0.4, gamma=2, F1-optimized threshold) and affect reported neural-net scores.
  • SMOTE oversampling of authentication class = SMOTE applied (params not fully specified)
    Synthetic balancing choice that changes the training distribution relative to natural stage durations.
axioms (5)
  • domain assumption The interleaved QIA protocol of Shaban & Ismail [19] is the right object of study: authentication rounds are secretly scheduled by a pre-shared key among data slots.
    Sec. II adopts [19] as the protocol under test; stage labels and security story depend on that design.
  • domain assumption A passive beam-splitter tap plus SPAD/power-meter observations does not constitute a quantum measurement that collapses the communicated states in a way that breaks the protocol’s CHSH/entanglement signature.
    Threat model Sec. III.B and validation Sec. V.A (S≈2.32–2.36 > 2).
  • domain assumption Stage-conditioned differences in polarization preparation and timing produce stable, learnable classical signatures in count rate, inter-arrival statistics, and optical power.
    Implicit throughout feature design (Sec. IV.C–D) and required for any generalization beyond this apparatus.
  • standard math Standard supervised classifiers (RF/LSTM/GRU) with engineered features are appropriate estimators of the map f: X→S.
    Sec. IV.E; ordinary ML methodology.
  • ad hoc to paper Experimenter-assigned segment labels after regrouping continuous traces into data-then-auth blocks correctly represent protocol stages for training and test.
    Sec. IV.B preprocessing step that creates the supervised dataset; not available to a field attacker a priori.

pith-pipeline@v1.2.0-grok45-kimik3 · 17425 in / 3666 out tokens · 81016 ms · 2026-07-31T09:26:18.067900+00:00 · methodology

0 comments
read the original abstract

Quantum networks can enable distributed computing and sensing. To realize these capabilities securely, quantum identity authentication is essential. Without authentication at the quantum layer, malicious repeaters may retain entanglement instead of performing swapping, enabling man-in-the-middle attacks (MitM) between communicating parties. Authentication mitigates this threat by embedding authentication qubits within data qubits at positions and bases based on a secret key shared a priori. While prior work analyzes security and MitM detection guarantees, physical layer side channel analysis remains unexplored. If an attacker infers protocol stages, it can avoid authentication qubits and extract data qubits, rendering authentication ineffective. To this end, we carry out experimental studies using a quantum communication testbed. A beam splitter is used to tap a portion of the optical signal, allowing the observer to collect side channel data without disrupting the quantum state. We evaluate two sampling settings, where 30% or 10% of the signal is diverted. The collected side channel data includes photon arrival timing and optical power data obtained using a single-photon detector and a power meter. Using this dataset, we extract and engineer features that capture both timing dynamics and signal intensity variations. We then train machine learning models to classify protocol stages based solely on side channel observations. Our results show that protocol-stage inference is feasible with high accuracy, reaching 98% (F1-score 97%) at 30% sampling and 96% (F1-score 94%) at 10% sampling. These findings reveal an overlooked vulnerability and highlight the need for robust designs against side channel inference attacks.

Figures

Figures reproduced from arXiv: 2607.24639 by Blaine Keyton, Carrie Houston, Contessa Wilburn, Lance Young, Marwan Elawady, Mohamed Shaban, Muhammad Ismail.

Figure 1
Figure 1. Figure 1: (1) laser driver, (2) 405 nm pump laser source, (3) iris, (4) kinematic mirrors for beam alignment, (5) alignment laser, [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Effect of passive sampling on signal photon count [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Confusion matrices illustrating the impact of feature modality and sampling ratio on protocol classification. (a) power [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Feature importance for the power meter dataset: (a) [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Feature importance for the time tagger dataset: (a) [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Feature importance for the combined dataset: (a) [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] 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

24 extracted references · 1 canonical work pages

  1. [1]

    Quantum Internet: A vision for the road ahead,

    S. Wehner, D. Elkouss, and R. Hanson, “Quantum Internet: A vision for the road ahead,”Science, vol. 362, no. 6412, p. eaam9288, 2018

    2018

  2. [2]

    Secure quantum key distribution with realistic devices,

    F. Xu, X. Ma, Q. Zhang, H.-K. Lo, and J.-W. Pan, “Secure quantum key distribution with realistic devices,”Reviews of modern physics, vol. 92, no. 2, p. 025002, 2020

    2020

  3. [3]

    A short review on quantum identity authenti- cation protocols: how would Bob know that he is talking with Alice?

    A. Dutta and A. Pathak, “A short review on quantum identity authenti- cation protocols: how would Bob know that he is talking with Alice?” Quantum Information Processing, vol. 21, no. 11, p. 369, 2022

    2022

  4. [4]

    Single trace side channel analysis on quantum key distribution,

    S. Kim, S. Jin, Y . Lee, B. Park, H. Kim, and S. Hong, “Single trace side channel analysis on quantum key distribution,” in2018 International Conference on Information and Communication Technology Conver- gence (ICTC), 2018, pp. 736–739

    2018

  5. [5]

    Deep-learning-based radio-frequency side-channel attack on quantum key distribution,

    A. Baliuka, M. St ¨ocker, M. Auer, P. Freiwang, H. Weinfurter, and L. Knips, “Deep-learning-based radio-frequency side-channel attack on quantum key distribution,”Physical Review Applied, vol. 20, no. 5, Nov. 2023. [Online]. Available: http://dx.doi.org/10.1103/PhysRevAppl ied.20.054040

    work page doi:10.1103/physrevappl 2023

  6. [6]

    Experimental side channel analysis of BB84 QKD source,

    A. Biswas, A. Banerji, P. Chandravanshi, R. Kumar, and R. P. Singh, “Experimental side channel analysis of BB84 QKD source,”IEEE Journal of Quantum Electronics, vol. 57, no. 6, pp. 1–7, 2021

    2021

  7. [7]

    Quantum circuit reconstruction from power side-channel attacks on quantum computer controllers,

    F. Erata, C. Xu, R. Piskac, and J. Szefer, “Quantum circuit reconstruction from power side-channel attacks on quantum computer controllers,”IACR Transactions on Cryptographic Hardware and Embedded Systems, vol. 2024, no. 2, p. 735–768, Mar. 2024. [Online]. Available: http://dx.doi.org/10.46586/tches.v2024.i2.735-768

    doi:10.46586/tches.v2024.i2.735-768 2024

  8. [8]

    Quantum leak: Timing side-channel attacks on cloud-based quantum services,

    C. Lu, E. Telang, A. Aysu, and K. Basu, “Quantum leak: Timing side-channel attacks on cloud-based quantum services,” inProceedings of the Great Lakes Symposium on VLSI 2025, ser. GLSVLSI ’25. New York, NY , USA: Association for Computing Machinery, 2025, p. 252–257. [Online]. Available: https://doi.org/10.1145/3716368.3735264

    arXiv 2025

  9. [9]

    Qtime: A machine learning framework for timing side-channel analysis in quantum circuit simulators,

    B. Dong, H. Feng, and Q. Wang, “Qtime: A machine learning framework for timing side-channel analysis in quantum circuit simulators,” in2025 IEEE 43rd International Conference on Computer Design (ICCD), 2025, pp. 335–341

    2025

  10. [10]

    Semi-quantum mutual identity authentication using bell states,

    S. Jiang, R.-G. Zhou, and W. Hu, “Semi-quantum mutual identity authentication using bell states,”International Journal of Theoretical Physics, vol. 60, no. 9, pp. 3353–3362, Sep 2021

    2021

  11. [11]

    An efficient quantum identity au- thentication key agreement protocol without entanglement,

    H. Zhu, L. Wang, and Y . Zhang, “An efficient quantum identity au- thentication key agreement protocol without entanglement,”Quantum Information Processing, vol. 19, no. 10, p. 381, Oct 2020

    2020

  12. [12]

    Quantum identity authentication in the counterfactual quantum key distribution protocol,

    B. Liuet al., “Quantum identity authentication in the counterfactual quantum key distribution protocol,”Entropy, vol. 21, no. 5, 2019

    2019

  13. [13]

    Controlled quantum secure direct commu- nication with authentication protocol based on five-particle cluster state and classical xor operation,

    X.-y. Zheng and Y .-x. Long, “Controlled quantum secure direct commu- nication with authentication protocol based on five-particle cluster state and classical xor operation,”Quantum Information Processing, vol. 18, no. 5, p. 129, Mar 2019

    2019

  14. [14]

    Quantum identity authentication scheme of vehicular ad-hoc networks,

    Z. Chen, K. Zhou, and Q. Liao, “Quantum identity authentication scheme of vehicular ad-hoc networks,”International Journal of The- oretical Physics, vol. 58, no. 1, pp. 40–57, Jan 2019

    2019

  15. [15]

    Quantum identity authentication with single photon,

    C. h. Hong, J. Heo, J. G. Jang, and D. Kwon, “Quantum identity authentication with single photon,”Quantum Information Processing, vol. 16, no. 10, p. 236, Aug 2017

    2017

  16. [16]

    Quantum identity au- thentication based on the extension of quantum rotation,

    G. Chen, Y . Wang, L. Jian, Y . Zhou, and S. Liu, “Quantum identity au- thentication based on the extension of quantum rotation,”EPJ Quantum Technology, vol. 10, no. 1, p. 11, Apr 2023

    2023

  17. [17]

    Controlled secure direct quantum communi- cation inspired scheme for quantum identity authentication,

    A. Dutta and A. Pathak, “Controlled secure direct quantum communi- cation inspired scheme for quantum identity authentication,”Quantum Information Processing, vol. 22, no. 1, p. 13, Dec 2022

    2022

  18. [18]

    A novel quantum identity authentication protocol without entanglement and preserving pre-shared key informa- tion,

    B. D. Rao and R. Jayaraman, “A novel quantum identity authentication protocol without entanglement and preserving pre-shared key informa- tion,”Quantum Information Processing, vol. 22, no. 2, p. 92, Jan 2023

    2023

  19. [19]

    Secured quantum identity authentication pro- tocol for quantum networks,

    M. Shaban and M. Ismail, “Secured quantum identity authentication pro- tocol for quantum networks,” in2024 IEEE 100th Vehicular Technology Conference (VTC2024-Fall), 2024, pp. 1–6

    2024

  20. [20]

    SMOTE: synthetic minority over-sampling technique,

    N. V . Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, “SMOTE: synthetic minority over-sampling technique,”Journal of arti- ficial intelligence research, vol. 16, pp. 321–357, 2002

    2002

  21. [21]

    Random forests,

    L. Breiman, “Random forests,”Machine learning, vol. 45, no. 1, pp. 5–32, 2001

    2001

  22. [22]

    Long short-term memory,

    S. Hochreiter and J. Schmidhuber, “Long short-term memory,”Neural computation, vol. 9, no. 8, pp. 1735–1780, 1997

    1997

  23. [23]

    Empirical evaluation of gated recurrent neural networks on sequence modeling,

    J. Chung, C. Gulcehre, K. Cho, and Y . Bengio, “Empirical evaluation of gated recurrent neural networks on sequence modeling,”arXiv preprint arXiv:1412.3555, 2014

    Pith/arXiv arXiv 2014

  24. [24]

    Permutation importance: a corrected feature importance measure,

    A. Altmann, L. Tolos ¸i, O. Sander, and T. Lengauer, “Permutation importance: a corrected feature importance measure,”Bioinformatics, vol. 26, no. 10, pp. 1340–1347, 2010

    2010