Pith. sign in

REVIEW 4 major objections 4 minor 51 references

LLM-Assisted Cheating Detection in Korean Language via Keystrokes

T0 review · 4 major / 4 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 Keystroke timing can distinguish Korean essays written from scratch, paraphrased from ChatGPT, or transcribed from ChatGPT, with accuracy up to about 93 percent.

desk verdict Useful new Korean keystroke dataset for LLM-cheating detection, but the fixed scenario order threatens the main comparison and the human-baseline 'significance' isn't tested. read the letter →

arxiv 2507.22956 v1 pith:QHLRXYFK submitted 2025-07-29 cs.LG cs.HC

Dong Hyun Roh , Rajesh Kumar , An Ngo This is my paper · ORCID
classification cs.LGcs.HC
keywords keystrokedynamicsLLM-assistedcheatingacademicintegrityKoreanlanguageBloom'staxonomycognitiveloadparaphrasedetectiontemporalandrhythmicfeatures
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

This paper tries to establish that keystroke dynamics—the timing and rhythm of how a person types—can reliably reveal whether a Korean-language text was written honestly, paraphrased from a ChatGPT response, or transcribed from one. The authors built the first Korean keystroke dataset for LLM-assisted cheating, with 69 participants writing under three conditions across six levels of Bloom's Taxonomy, and trained standard classifiers on temporal and rhythmic features. In user-independent tests the models reached roughly 85–93 percent accuracy, with transcribed responses easiest to detect and paraphrased responses hardest. The same models clearly outperformed human raters, who identified paraphrased responses only about 42 percent of the time. If the claim holds, keystroke monitoring could serve as a scalable, low-intrusion complement to text-based plagiarism tools in non-English academic settings.

What carries the argument

The load-bearing machinery is the pairing of two interpretable keystroke feature sets with a cognitive-load-aware evaluation design. Temporal features—Key Hold Time (KHT) and Key Interval Time (KIT) summarized by quartiles, mean, and standard deviation—capture motor fluency and physical execution. Rhythmic features capture cognitively driven planning and revision: fifteen types including nine binned pause durations, P-bursts (text production segments ended by a pause of at least two seconds), R-bursts (writing sequences ending in deletion), delete bursts, inter-word and inter-sentence pauses, and pause-before-delete, summarized by seven statistics plus entropy, yielding 107 features. The evaluation splits prompts into low-load (Bloom's remember, understand, apply) and high-load (analyze, evaluate, create) groups, and compares Cognition-Unaware, Cognition-Aware (H→H, L→L), and cross-cognition (H→L, L→H) settings using user-independent splits of the 69 participants and MLP, SVM, and XGBoost classifiers with genetic-algorithm hyperparameter tuning.

What would settle it

Run the identical three-condition protocol with participants randomly assigned to different scenario orders and compare per-class accuracy; if accuracy drops or the temporal-feature advantage disappears when order is controlled, the central claim is an artifact of session ordering. A simpler check is to add elapsed time since session start as a feature and see whether it alone predicts the labels.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the way a person types carries a reliable behavioral signature of how the text was produced, even when the text itself reads fluently. Using Key Hold Times and Key Interval Times as temporal features, and binned pauses, bursts, and deletion patterns as rhythmic features, the authors report that classifiers distinguish bona fide Korean writing from paraphrased and transcribed ChatGPT responses under both Cognition-Unaware and Cognition-Aware splits. Temporal features perform best when training and test tasks share the same cognitive load, while rhythmic features generalize better when cognitive load changes across train and test. Detecting transcribed responses was consistently easiest, with recall near or above 94 percent; paraphrased responses were the most confusable, especially with bona fide writing. A human baseline rated by 18 fluent Korean speakers was far weaker, correctly identifying only 69.44 percent of bona fide, 41.67 percent of paraphrased, and 49.07 percent of transcribed responses, which the paper takes as evidence that keystroke-based behavioral detection adds information beyond what text inspection provides.

Load-bearing premise

The three writing conditions were performed in a fixed order without reported counterbalancing, so condition is entangled with session time, fatigue, and practice; if that is true, the classifiers may be detecting time-on-task rather than the writing strategy.

Editorial extensions

If this is right

  • In online Korean assessments, keystroke logging could flag suspected ChatGPT paraphrasing for human review, catching a form of cheating that text-matching tools miss.
  • Transcribed ChatGPT text is the easiest to detect, so proctoring systems can prioritize high-confidence transcription flags with low false-positive risk.
  • Because temporal features excel when cognitive load is known and rhythmic features excel when it varies, a fusion of both feature sets would likely outperform either alone.
  • User-independent detection at 85–93 percent accuracy suggests the approach does not require enrolling a user's typing profile in advance.
  • The public dataset gives other researchers a Korean-language benchmark for keystroke-based authorship and academic-integrity studies.

Reading between the lines

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

  • If the three writing scenarios were always administered in the same order, part of the reported accuracy may reflect elapsed session time, fatigue, or practice rather than the cognitive writing strategy; a randomized-order replication would show how much of the signal is truly behavioral.
  • The rhythmic features' robustness across cognitive-load transfer hints that pause and burst statistics approximate stable motor-cognitive traits, so similar detectors might transfer to other languages or to typing on different keyboards—an extension the paper does not test.
  • The finding that human raters struggle most with paraphrase suggests that LLM-detection systems relying only on text content will keep failing; combining keystroke behavior with text-level markers is a concrete next step.
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 / 4 minor

Summary. The paper introduces a Korean keystroke dataset from 69 participants who completed writing tasks under three conditions: bona fide writing, paraphrasing ChatGPT responses, and transcribing ChatGPT responses, with six prompts per condition aligned to Bloom's Taxonomy. It extracts interpretable temporal and rhythmic keystroke features and evaluates MLP, SVM, and XGBoost classifiers under user-independent splits in Cognition-Unaware and Cognition-Aware settings. The authors report roughly 85–93% classification accuracy, with transcribed responses easiest to detect, paraphrased responses hardest, temporal features stronger under Cognition-Aware settings, rhythmic features more robust under cross-cognition transfer, and models outperforming human raters. The dataset and code are publicly available.

Significance. If the results hold, this is a useful contribution: it provides the first Korean-language keystroke dataset for LLM-assisted cheating detection, uses a finer-grained three-class labeling of cheating behavior, and evaluates generalization to unseen users. The internal numbers are broadly consistent with the main claim, and the evaluation is not circular: accuracies are computed on held-out users, and features are not fitted to the labels. The public dataset and code are a concrete reproducibility asset. The main risks are experimental-design confounds and missing statistical support for headline comparisons, rather than methodological circularity.

major comments (4)
  1. [Section 3.1 and 3.2] The paper never states that the order of the three writing scenarios was randomized or counterbalanced. Section 3.2 reports that only Phase 1 data are used and Section 3.1 describes a five-minute break between scenarios, but if the scenario order was fixed, then the scenario label is confounded with session position, fatigue, and practice with the data-collection portal. Every accuracy reported in Figures 2–6 could then partly reflect a classifier reading the session clock or interface familiarity rather than the cognitive writing strategy. Because the dataset and code are public, please either state the randomization/counterbalancing protocol or run an order-effect check (e.g., compare feature distributions or classifier performance by session position within each scenario) and report the result.
  2. [Section 4.4 and Abstract] The abstract and Section 4.4 claim that the models 'significantly outperform' human evaluators, but no significance test is reported. The human baseline uses only 18 raters and 18 responses, provides no chance-level benchmark or confidence intervals, and compares text-only human judgments with keystroke-based model predictions, which are different information modalities. Please add a proper statistical comparison (e.g., bootstrap or a mixed-effects model with rater and item random effects), report inter-rater agreement, and qualify the claim if the comparison is not statistically supported.
  3. [Section 3.2] The central claim depends on the scenario labels being correct, but the paper does not quantify adherence. Section 3.2 lists three validation steps, including verifying word counts and checking paraphrase quality, yet it reports no statistics from these checks (e.g., how many ChatGPT responses had to be re-prompted, how many participant responses were rejected, or how much paraphrases differed textually from their ChatGPT source). Please report adherence statistics or a text-similarity analysis between paraphrased responses and their source; otherwise label noise is unmeasured and the reported accuracies may be optimistic.
  4. [Section 3.5] The Cognition-Aware analysis defines high and low cognitive load by question index (questions 1–3 vs 4–6) without any independent validation of the load manipulation. The Section 4.2 and 4.3 claims about high-load discriminability and cross-cognition transfer rest entirely on this assignment. Please justify the mapping with pilot ratings, response-time data, or a manipulation check, or relabel the axis as 'question set' rather than cognitive load.
minor comments (4)
  1. [Section 2] There is a typographical issue in 'difficult to fake ,' with a stray space before the comma; please proofread the manuscript.
  2. [Section 4.4] The phrase 'minimal contextual guidance and explicit definitions of each label were provided' is internally contradictory; please clarify exactly what guidance and definitions the human raters received.
  3. [Section 3.6 and Figures 2, 4, 6] The text refers to confidence bands and varying training sizes, but the details of how the confidence bands were computed are not specified; please state the estimator and the number of repetitions used.
  4. [Section 3.4.2] The sentence listing the seven statistics reads as if 'first and third quartiles' is a single item; please rephrase as 'first quartile, third quartile' to avoid confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claim is supported by held-out user evaluation, and no fitted parameter is reused as a prediction.

full rationale

I walked the claimed derivation chain: scenario labels are set by the experimental protocol; keystroke features are extracted from raw event timestamps; classifiers are evaluated user-independently on non-overlapping train/test user splits; feature selection and genetic-algorithm hyperparameter tuning are confined to training splits. The reported accuracies are therefore not constructed from the labels or from test-set information. The self-citations present in the paper (e.g., Kundu et al. [2], Kumar et al. [24,25,28], Khurana et al. [44]) are used for background, feature conventions, or planned future methodology, and none carries the load of the central transfer claim. The possible session-order confound identified during review is a validity or experimental-design concern, not a circular reduction: it does not make the derivation equivalent to its inputs by definition. No equation, fitted parameter, or renamed known result is reused as the predicted outcome, so no circular step can be exhibited.

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

The central empirical claims rest on four main assumptions: keystroke features reflect cognitive states, the Bloom's Taxonomy split creates meaningful load differences, participants actually followed the assigned writing conditions, and sample independence after user splitting. None of these is formally validated; the dataset and code are public, which allows downstream checks, but within the paper they are taken as given.

free parameters (4)
  • Pause bin width = 300 ms
    Nine bins from 3 ms to 2400 ms+; hand-chosen, no sensitivity analysis reported.
  • P-burst pause threshold = 2 s
    Defines P-burst boundaries; based on prior work but no validation on these data.
  • Feature selection ratio = 50%
    Top half of features by mutual information kept per split; chosen without ablation.
  • Classifier hyperparameters = Not reported
    Tuned per split via genetic algorithm with accuracy fitness; final values absent, hurting reproducibility.
assumptions (4)
  • domain assumption Keystroke behavioral features are valid proxies for cognitive processes such as planning, revision, and evaluation.
    Section 3.4.2 grounds the rhythmic features in Hayes' model and prior keystroke writing research, but no independent check that these features actually measure the intended cognitive states in Korean.
  • domain assumption Bloom's Taxonomy questions classified as low (1-3) and high (4-6) induce distinct cognitive load.
    Section 3.5 assigns low/high labels by question order, with no manipulation check, self-report, or behavioral validation of cognitive load differences.
  • domain assumption Participants complied with the assigned writing condition and manual validation captured noncompliance.
    Section 3.2 describes researcher verification, but a participant could still mentally compose a ChatGPT response from memory during the bona fide condition; the dataset has no independent ground truth beyond self-report.
  • standard math Multiple responses from the same participant are treated as independent samples after user-level splitting.
    The evaluation splits users, not samples, yet accuracy is computed over all question-level responses; repeated-measures correlation within a user is not accounted for, which can inflate perceived model quality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LLM-Assisted Cheating Detection in Korean Language via Keystrokes." pith.science (2026). https://pith.science/paper/QHLRXYFK

@misc{pith2026250722956,
  author       = {Pith},
  title        = {Pith review of: LLM-Assisted Cheating Detection in Korean Language via Keystrokes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QHLRXYFK}},
  note         = {Machine review of arXiv:2507.22956}
}
read the original abstract

This paper presents a keystroke-based framework for detecting LLM-assisted cheating in Korean, addressing key gaps in prior research regarding language coverage, cognitive context, and the granularity of LLM involvement. Our proposed dataset includes 69 participants who completed writing tasks under three conditions: Bona fide writing, paraphrasing ChatGPT responses, and transcribing ChatGPT responses. Each task spans six cognitive processes defined in Bloom's Taxonomy (remember, understand, apply, analyze, evaluate, and create). We extract interpretable temporal and rhythmic features and evaluate multiple classifiers under both Cognition-Aware and Cognition-Unaware settings. Temporal features perform well under Cognition-Aware evaluation scenarios, while rhythmic features generalize better under cross-cognition scenarios. Moreover, detecting bona fide and transcribed responses was easier than paraphrased ones for both the proposed models and human evaluators, with the models significantly outperforming the humans. Our findings affirm that keystroke dynamics facilitate reliable detection of LLM-assisted writing across varying cognitive demands and writing strategies, including paraphrasing and transcribing LLM-generated responses.

Figures

Figures reproduced from arXiv: 2507.22956 by the authors.

Figure 1
Figure 1. Overview of the data collection setup. Figure [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Performance of Temporal and Rhythmic feature-based models under Cognition-Unaware conditions. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Class-wise performance under Cognition-Unaware conditions. Temporal features in Teal blue, and Rhythmic features in Orange. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Performance of Temporal and Rhythmic feature-based models under Cognition-Aware (H [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Class-wise performance under Cognition-Aware conditions. Temporal features in Teal blue, and Rhythmic features in Orange. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Performance of Temporal and Rhythmic feature-based [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 51 canonical work pages

  1. [1]

    Crossley, Y

    S. Crossley, Y . Tian, J. Choi, L. Holmes, and W. Morris. Plagiarism detection using keystroke logs. In IEEE ICEDM,

    work page

  2. [2]

    Kundu, A

    D. Kundu, A. Mehta, R. Kumar, N. Lal, A. Anand, A. Singh, and R. Shah. Keystroke dynamics against academic dishonesty in the age of llms. IEEE IJCB, 2024. 1, 2, 4

    work page 2024

  3. [3]

    M. Hoq, Y . Shi, J. Leinonen, Damilola B., Collin L., T. Price, and B. Akram. Detecting chatgpt-generated code submissions in a cs1 course using machine learning models. SIGCSE, 2024. 1

    work page 2024

  4. [4]

    Using keystroke dynamics to detect nonoriginal text

    Paul Deane, Mo Zhang, Jiangang Hao, and Chen Li. Using keystroke dynamics to detect nonoriginal text. Journal of Educational Measurement, 2025. 1, 5, 6

    work page 2025

  5. [5]

    Keystroke dynamics as signal for shallow syntactic parsing

    Barbara Plank. Keystroke dynamics as signal for shallow syntactic parsing. In COLING, Osaka, Japan, 2016. The COLING 2016 Organizing Committee. 1, 2

    work page 2016

  6. [6]

    Agarwal, N

    N. Agarwal, N. Danielsen, P. Gravdal, and P. Bours. Contract cheat detection using biometric keystroke dynamics. In ICETA, 2022. 1

    work page 2022

  7. [7]

    Elliott Casal and Matt Kessler

    J. Elliott Casal and Matt Kessler. Can linguists distinguish between chatgpt/ai and human writing?: A study of research ethics and academic publishing. Research Methods in Applied Linguistics, 2023. 1, 2

    work page 2023

  8. [8]

    A large-scale comparison of human-written versus chatgpt-generated essays

    Steffen Herbold, Annette Hautli-Janisz, Ute Heuer, Zlata Kikteva, and Alexander Trautsch. A large-scale comparison of human-written versus chatgpt-generated essays. Scientific Reports, 2023. 1

    work page 2023

Show all 51 references
  1. [9]

    E-cheating prevention measures: detection of cheating at online examinations using deep learning approach–a case study

    Leslie Ching Ow Tiong and HeeJeong Jasmine Lee. E-cheating prevention measures: detection of cheating at online examinations using deep learning approach–a case study. https://arxiv.org/abs/2101.09841, 2021. 1

    2021 arXiv

  2. [10]

    Automated cheating detection based on video surveillance in the examination classes

    M Roa’a, Ibtisam A Aljazaery, and AHM Alaidi. Automated cheating detection based on video surveillance in the examination classes. iJIM, 2022. 1

    2022

  3. [11]

    Student cheating detection in higher education by implementing machine learning and lstm techniques

    Waleed Alsabhan. Student cheating detection in higher education by implementing machine learning and lstm techniques. Sensors, 2023. 1

    2023

  4. [12]

    Machine learning-based approach to exam cheating detection

    Firuz Kamalov, Hana Sulieman, and David Santandreu Calonge. Machine learning-based approach to exam cheating detection. Plos one, 2021. 1

    2021

  5. [13]

    Keystroke analysis: Reflections on procedures and measures

    Veerle M Baaijen, David Galbraith, and Kees De Glopper. Keystroke analysis: Reflections on procedures and measures. Written Communication, 2012. 1

    2012

  6. [14]

    Using keystroke analytics to understand cognitive processes during writing

    Mo Zhang, Hongwen Guo, and Xiang Liu. Using keystroke analytics to understand cognitive processes during writing. IEDMS, 2021. 1, 2

    2021

  7. [15]

    A cognitive process theory of writing

    Linda Flower and John R Hayes. A cognitive process theory of writing. College Composition and Communication, 1981. 1, 2

    1981

  8. [16]

    Modeling and remodeling writing

    John R Hayes. Modeling and remodeling writing. Written communication, 2012. 1, 2, 5

    2012

  9. [17]

    From keystrokes to cognitive processes: Analyzing morphological knowledge using keystroke logging

    Andreas Sellstone. From keystrokes to cognitive processes: Analyzing morphological knowledge using keystroke logging. Master’s thesis, Ume ˚a University, 2023. 1, 2, 5

    2023

  10. [18]

    Investigating cognitive rhythms as a new modality for continuous authentication

    Kiran S Balagani. Investigating cognitive rhythms as a new modality for continuous authentication. Technical report, NYIT, 2013. 1, 2, 4, 5

    2013

  11. [19]

    Detecting deception by manipulating cognitive load

    Aldert Vrij, Ronald Fisher, Samantha Mann, and Sharon Leal. Detecting deception by manipulating cognitive load. Trends in cognitive sciences, 2006. 1

    2006

  12. [20]

    Anderson and David R

    Lorin W. Anderson and David R. Krathwohl, editors. A Taxonomy for Learning, Teaching, and Assessing. A Revision of Bloom’s Taxonomy of Educational Objectives . Allyn & Bacon, 2001. 1, 2, 3

    2001

  13. [21]

    Identity authentication based on keystroke latencies

    Rick Joyce and Gopal Gupta. Identity authentication based on keystroke latencies. Commun. ACM, 1990. 2

    1990

  14. [22]

    Verifying identity via keystroke characterstics

    John Leggett and Glen Williams. Verifying identity via keystroke characterstics. International Journal of Man-Machine Studies, 1988. 2

    1988

  15. [23]

    Fabian Monrose and Aviel D. Rubin. Keystroke dynamics as a biometric for authentication. FGCS, 2000. 2

    2000

  16. [24]

    Phoha, and Abdul Serwadda

    Rajesh Kumar, Vir V . Phoha, and Abdul Serwadda. Continuous authentication of smartphone users by fusing typing, swiping, and phone movement patterns. In IEEE BTAS, 2016. 2

    2016

  17. [25]

    Rajesh Kumar, Partha Pratim Kundu, and Vir V . Phoha. Continuous authentication using one-class classifiers and their fusion. In IEEE ISBA, 2018. 2

    2018

  18. [26]

    On continuous user authentication via typing behavior

    Joseph Roth, Xiaoming Liu, and Dimitris Metaxas. On continuous user authentication via typing behavior. IEEE Transactions on Image Processing, 2014. 2

    2014

  19. [27]

    Investigating the discriminative power of keystroke sound

    Joseph Roth, Xiaoming Liu, Arun Ross, and Dimitris Metaxas. Investigating the discriminative power of keystroke sound. IEEE TIFS, 2015. 2

    2015

  20. [28]

    Spotting fake profiles in social networks via keystroke dynamics

    Alvin Kuruvilla, Rojanaye Daley, and Rajesh Kumar. Spotting fake profiles in social networks via keystroke dynamics. In IEEE-CCNC, 2024. 2, 5

    2024

  21. [29]

    Ieee-bigmm

    Vishaal Udandarao, Mohit Agrawal, Rajesh Kumar, and Rajiv Ratn Shah. Ieee-bigmm. In IEEE-BigMM, 2020. 2

    2020

  22. [30]

    Monaco, Ruben Vera-Rodriguez, and Julian Fierrez

    Alejandro Acien, Aythami Morales, John V . Monaco, Ruben Vera-Rodriguez, and Julian Fierrez. Typenet: Deep learning keystroke biometrics. IEEE TBIOM, 2022. 2, 4

    2022

  23. [31]

    Plagiarism deterrence in cs1 through keystroke data

    Kaden Hart, Chad Mano, and John Edwards. Plagiarism deterrence in cs1 through keystroke data. In SIG-CSE, 2023. 2

    2023

  24. [32]

    Shared multi-keyboard and bilingual datasets to support keystroke dynamics research

    Ahmed Anu Wahab, Daqing Hou, Mahesh Banavar, Stephanie Schuckers, Kenneth Eaton, Jacob Baldwin, and Robert Wright. Shared multi-keyboard and bilingual datasets to support keystroke dynamics research. InACM CODASPY,

  25. [33]

    Alsuhibany, Muna Almushyti, Noorah Alghasham, and Fatimah Alkhudier

    Suliman A. Alsuhibany, Muna Almushyti, Noorah Alghasham, and Fatimah Alkhudier. Analysis of free-text keystroke dynamics for arabic language using euclidean distance. In 12th International Conference on Innovations in Information Technology (IIT), 2016. 2

    2016

  26. [34]

    Keystroke timing analysis for individual identification in japanese free text typing

    Toshiharu Samura and Haruhiko Nishimura. Keystroke timing analysis for individual identification in japanese free text typing. In ICCAS-SICE, 2009. 2

    2009

  27. [35]

    Language dependent challenge-based keystroke dynamics

    Patrick Bours and Sukanya Brahmanpally. Language dependent challenge-based keystroke dynamics. In IEEE ICCST, 2017. 2

    2017

  28. [36]

    User-representative feature selection for keystroke dynamics

    Eesa Al Solami, Colin Boyd, Andrew Clark, and Irfan Ahmed. User-representative feature selection for keystroke dynamics. In 5th International Conference on Network and System Security, 2011. 2

    2011

  29. [37]

    Shared data set for free-text keystroke dynamics authentication algorithms

    Catalin Iapa and Vladimir Cretu. Shared data set for free-text keystroke dynamics authentication algorithms. In Preprint,

  30. [38]

    On continuous user authentication via hidden free-text based monitoring

    Elena Kochegurova, Elena Luneva, and Ekaterina Gorokhova. On continuous user authentication via hidden free-text based monitoring. In IITI. Springer International Publishing, 2019. 2

    2019

  31. [39]

    Freely typed keystroke dynamics-based user authentication for mobile devices based on heterogeneous features

    Junhong Kim and Pilsung Kang. Freely typed keystroke dynamics-based user authentication for mobile devices based on heterogeneous features. Pattern Recognition, 2020. 2

    2020

  32. [40]

    Stragapede

    Giuseppe et al. Stragapede. Ieee bigdata 2023 keystroke verification challenge (kvc). In 2023 IEEE International Conference on Big Data (BigData), 2023. 4

    2023

  33. [41]

    Typeformer: Transformers for mobile keystroke biometrics

    Giuseppe Stragapede, Paula Delgado-Santos, Ruben Tolosana, Ruben Vera-Rodriguez, Richard Guest, and Aythami Morales. Typeformer: Transformers for mobile keystroke biometrics. Neural Computing and Applications ,

  34. [42]

    Wahab and D

    A. Wahab and D. Hou. When simple statistical algorithms outperform deep learning: A case of keystroke dynamics. In ICPRAM, 2023. 4

    2023

  35. [43]

    Survey of review spam detection using machine learning techniques

    Michael Crawford, Taghi M Khoshgoftaar, Joseph D Prusa, Aaron N Richter, and Hamzah Al Najada. Survey of review spam detection using machine learning techniques. Journal of Big Data, 2015. 4

    2015

  36. [44]

    Synthesizing human gaze feedback for improved NLP performance

    Varun Khurana, Yaman Kumar, Nora Hollenstein, Rajesh Kumar, and Balaji Krishnamurthy. Synthesizing human gaze feedback for improved NLP performance. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics (EACL), 2023. 5

    2023

  37. [45]

    Alejandro et al. Acien. Detection of mental fatigue in the general population: Feasibility study of keystroke dynamics as a real-world biomarker. JMIR Biomedical Engineering ,

  38. [46]

    Keystroke dynamics: Concepts, techniques, and applications

    Rashik Shadman, Ahmed Anu Wahab, Michael Manno, Matthew Lukaszewski, Daqing Hou, and Faraz Hussain. Keystroke dynamics: Concepts, techniques, and applications. ACM Computing Survey, 2025. 5

    2025

  39. [47]

    Examining Pauses in Writing: Theory, Methods and Empirical Data

    Wengelin, ˚Asa. Examining Pauses in Writing: Theory, Methods and Empirical Data. Computer key-stroke logging and writing: methods and applications (Studies in Writing) ,

  40. [48]

    Keystroke patterns as prosody in digital writings: A case study with deceptive reviews and essays

    Ritwik Banerjee, Song Feng, Jun Seok Kang, and Yejin Choi. Keystroke patterns as prosody in digital writings: A case study with deceptive reviews and essays. In EMNLP, 2014. 5

    2014

  41. [49]

    Detecting contract cheating using learning analytics

    Kelly Trezise, Tracii Ryan, Paula de Barba, and Gregor Kennedy. Detecting contract cheating using learning analytics. Journal of Learning Analytics, 2019. 5

    2019

  42. [50]

    Using keystroke behavior patterns to detect nonauthentic texts in writing assessments: Evaluating the fairness of predictive models

    Yang Jiang, Mo Zhang, Jiangang Hao, Paul Deane, and Chen Li. Using keystroke behavior patterns to detect nonauthentic texts in writing assessments: Evaluating the fairness of predictive models. Journal of Educational Measurement ,

  43. [51]

    Young and Derek C

    Steven R. Young and Derek C. et al. Rose. Optimizing deep learning hyperparameters through an evolutionary algorithm. In MLHPC. ACM, 2015. 6

    2015

Pith tools

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