Pith. sign in

REVIEW 3 major objections 8 minor 49 references

This paper introduces SC-NLP-LMF, a six-phase NLP lifecycle protocol integrating security, bias audits, privacy, drift monitoring, and decommissioning, and claims it is the first compliance-aware NLP lifecycle framework.

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 · deepseek-v4-flash

2026-08-03 13:54 UTC pith:T26QQ4DN

load-bearing objection Useful synthesis of NLP lifecycle governance, but the 'practical applicability' claim rests on a hypothetical case study and an internal-access assumption that excludes black-box API deployments. the 3 major comments →

arxiv 2512.22060 v1 pith:T26QQ4DN submitted 2025-12-26 cs.CR cs.CLcs.CY

Toward Secure and Compliant AI: Organizational Standards and Protocols for NLP Model Lifecycle Management

Sunil Arora , John Hastings This is my paper
classification cs.CR cs.CLcs.CY
keywords NLP lifecycle governanceAI complianceEU AI Actbias auditingdrift detectionmodel decommissioningdifferential privacysecure model deployment
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.

Existing AI governance frameworks are too general to handle language-specific risks like semantic drift, evolving bias, and prompt-injection attacks; existing MLOps pipelines (e.g., TFX, Kubeflow) lack compliance and audit hooks. To fill that gap, the paper proposes SC-NLP-LMF, a six-phase lifecycle model — data governance, secure training, deployment governance, monitoring and drift detection, retraining and updates, and decommissioning/archival — synthesized from a PRISMA-guided review of 45 sources and mapped to NIST AI RMF, ISO/IEC 42001, the EU AI Act, and MITRE ATLAS. The framework assigns concrete practices to each phase (data statements, differential privacy, fairness toolkits, SHAP/LIME, model cards, TEVV-based monitoring, blue-green deployment, verifiable-claims archival). The paper claims this is the first framework with dedicated, phase-specific, compliance-aware structure for NLP lifecycle management, and illustrates the model with a narrative healthcare case study in which COVID-era terminology drift triggers retraining. A sympathetic reader would care because the framework offers a ready-made reference architecture for regulated sectors, though its reach depends on organizational visibility into model internals.

Core claim

The paper's central claim is that NLP systems need lifecycle-wide governance that is phase-specific and compliance-aware, and that such a structure can be assembled from existing practices and standards. Its SC-NLP-LMF organizes six phases — data governance, secure model training, deployment governance, monitoring and drift detection, retraining and updates, and decommissioning and archival — and binds each to concrete protocols, including differential privacy, Fairness Indicators/AI Fairness 360, SafeML, SHAP/LIME, model cards, TEVV, AI TRiSM, and verifiable-claims-based archival. The healthcare case study is intended to show that the framework detects terminology drift (new COVID-related t

What carries the argument

The load-bearing structure is the six-phase lifecycle itself, with each phase anchored to named protocols and standards. The monitoring-and-drift-detection phase acts as the trigger mechanism: when semantic drift or bias indicators fire, the framework routes the model through retraining-and-updates, followed by deployment governance and eventually decommissioning/archival. The PRISMA-based synthesis of 45 peer-reviewed and regulatory sources supplies the evidence base that maps each phase to existing tools, so the framework's novelty is the integration, not the invention, of controls.

Load-bearing premise

The framework assumes the organization has full visibility and control over model internals, training data, and evaluation pipelines; for SaaS, proprietary, or third-party API NLP deployments, most lifecycle phases cannot be directly exercised.

What would settle it

Run a one-year controlled pilot in a regulated NLP deployment: one team follows SC-NLP-LMF, another follows standard MLOps. If the framework group shows no measurable improvement in drift detection, bias mitigation, or audit-readiness, the claimed advantage is unsupported. Alternatively, attempt to apply Phase 2 (secure training) and Phase 6 (decommissioning) to a black-box third-party API model and observe that those phases are simply inapplicable without vendor cooperation.

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

If this is right

  • Regulated organizations get a single reference architecture that closes lifecycle gaps left by engineering-focused pipelines such as TFX and Kubeflow.
  • Continuous drift monitoring plus trigger-based retraining would give a concrete operational route to staying aligned with the EU AI Act and GDPR as language and regulations evolve.
  • Standardized phase practices (bias audits, differential privacy, model cards, decommissioning logs) would make internal and third-party audits more tractable.
  • Explicit decommissioning and archival protocols would prevent retired models from lingering in production, a neglected compliance and security risk.

Where Pith is reading between the lines

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

  • The framework's internal-visibility assumption makes it inapplicable to black-box APIs and SaaS LLMs as written; a natural extension is to adapt it into a contractual/audit overlay (documentation review, API behavior monitoring, third-party certifications) — but that is my inference, not the paper's claim.
  • The healthcare case study is narrative, not empirical; a falsifiable follow-up would run a real hospital NLP deployment with and without the framework to measure drift-detection latency, audit-readiness, and bias-mitigation outcomes.
  • The phase-trigger logic could be operationalized as a compliance-aware MLOps pipeline or automated governance agent, enabling quantified cost/benefit comparisons across SMEs and enterprises.

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

3 major / 8 minor

Summary. The paper proposes SC-NLP-LMF, a six-phase framework (Data Governance, Secure Model Training, Deployment Governance, Monitoring and Drift Detection, Retraining and Updates, Decommissioning and Archival) for governing NLP models throughout their lifecycle. The framework is derived from a PRISMA-style review of 45 sources and is mapped to existing standards including NIST AI RMF, ISO/IEC 42001:2023, the EU AI Act, and MITRE ATLAS. The authors claim that SC-NLP-LMF is the first NLP-specific, compliance-aware lifecycle framework and that it is practically applicable, based on a healthcare case study described as a narrative. The paper does not present empirical measurements, baselines, or external validation.

Significance. If substantiated, the framework could serve as a useful integration of existing governance, security, and fairness practices into an NLP-specific lifecycle, and its mapping to standards and toolkits (Fairlearn, AI Fairness 360, SHAP, LIME, SafeML, Model Cards, etc.) provides a structured reference for practitioners. The systematic selection of 45 sources and the explicit phase-tool mapping are strengths. However, the central claims of practical applicability and being 'the first' framework are not currently supported by the evidence: the only evaluation is a self-authored hypothetical case study, and the framework's internal-access assumption excludes a large and growing segment of real-world NLP deployments (SaaS, proprietary LLMs, third-party APIs). The contribution is best viewed at this stage as a conceptual framework proposal rather than a validated solution.

major comments (3)
  1. [V. Results; IV-A Healthcare Case Study] The abstract and Section VII state that evaluation across 'real-world use cases in healthcare and finance' demonstrates practical applicability. The only evaluation in the paper is the Section IV-A healthcare case study, which is explicitly a 'narrative outlines how each lifecycle phase would be enacted' in a hypothetical 'hospital network.' There are no data, no measurements, no comparison with existing frameworks, and no external validation; the finance use case is never described. Moreover, this narrative is authored by the same authors to illustrate their own framework, and is then cited in Section V as evidence that the framework can 'proactively guide organizations.' This is a circular validation path and does not support the load-bearing claim of practical applicability.
  2. [VI. Discussion] The framework assumes 'access to internal NLP model design and architectures, training data, and evaluation pipelines,' and the authors concede that this 'may not hold true for SaaS (Software as a Service), proprietary LLMs, or third-party APIs.' This is not a peripheral limitation: for black-box deployments, the proposed lifecycle phases — secure training, retraining, bias audits on training data, decommissioning — cannot be executed by the adopting organization. Since the paper claims lifecycle-wide applicability to high-risk domains, the assumption materially narrows the framework's reach. A concrete test would be to show how the 'contractual or auditing overlay' mentioned by the authors operates on an actual SaaS/API deployment, or to explicitly rescope the claims to white-box deployments only.
  3. [III. Methodology; V. Results] The paper claims that the framework was developed through a PRISMA-based systematic review, reporting 186 candidate documents and 45 selected works. However, no PRISMA flow diagram, search strings, database list, inclusion/exclusion criteria, or quality assessment are provided, so the review's reproducibility cannot be assessed. There is also an internal inconsistency: Section III says 45 documents were selected for synthesis, while Section V states that the framework was built by 'consolidating over 50 authoritative sources.' The authors should correct this and provide the PRISMA documentation if the systematic-review claim is to stand.
minor comments (8)
  1. [IV-A] SafeML is described as 'a proactive safety filter designed to detect vulnerability to adversarial examples and prompt-based misuse,' but the cited SafeML [30] is a privacy-preserving, Byzantine-robust distributed training framework, not an adversarial-robustness evaluator. The description should match the cited source.
  2. [V; VII] The finance use case mentioned in the abstract and conclusions is never described anywhere in the paper. Either add a concrete finance illustration or remove the reference to it.
  3. [Abstract; VII] The abstract says the framework 'ensures secure operation' and Section VII says it 'provides a defensible and actionable governance roadmap.' Given the limitations acknowledged in Section VI, these claims are too strong and should be qualified (e.g., 'supports' or 'guides' secure operation).
  4. [Table II] The SC-NLP-LMF row lists 'Integrated with Fairlearn, Audit Trails, secure deployment.' It would be clearer to distinguish tools that are directly integrated from those that are merely recommended or mapped to the phases.
  5. [II; Table I] The claim that SC-NLP-LMF is 'the first framework to offer a dedicated, phase-specific, and compliance-aware structure' (Section V) is not supported by a comprehensive novelty search; Table I compares only ten selected frameworks. Consider softening to 'one of the first' or providing a more systematic comparison.
  6. [References] Reference [20] ('Code of Practice for the Cyber Security of AI') and reference [11] ('AI Cyber Security Code of Practice') appear to describe the same UK guidance; if so, they should be consolidated or distinguished more clearly.
  7. [I] CVE-2025-8217 is cited through a blog post; including the official CVE/NVD or vendor advisory would improve reliability.
  8. [III] The PRISMA statement [21] is itself a reporting guideline; the authors should state which version (2009 vs 2020) they followed, and whether a protocol was registered.

Circularity Check

1 steps flagged

Practical-applicability claim rests on a self-authored narrative that walks through the framework's own phases; no independent evaluation is reported.

specific steps
  1. other [Section IV-A (Healthcare Case Study), Section V (Results), Section VII (Conclusions)]
    "The following narrative outlines how each lifecycle phase would be enacted under the proposed framework: ... The framework's effectiveness was further assessed by applying it to a detailed, narrative case study in healthcare NLP. ... Evaluation across real-world use cases in healthcare and finance demonstrates that the SC-NLP-LMF is not only theoretically sound but also practically applicable."

    The only reported validation is a hypothetical narrative written in Section IV-A to illustrate the framework's own six phases. Section V then treats that narrative as an 'effectiveness' assessment, and Section VII inflates it into 'real-world use cases in healthcare and finance' demonstrating practical applicability. The evidence is not external to the framework; it is the framework applied to itself. The central applicability claim thus reduces to a restatement of the self-authored illustration, and no independent benchmark, implementation, or measured outcome is provided.

full rationale

This paper does not derive quantitative predictions or fit parameters to data, so the classic circularity patterns (fitted input called prediction, self-citation chains, uniqueness theorems) are not present. The framework's content is synthesized from cited external standards, regulations, and toolkits, giving it independent grounding. However, the claim that the framework is practically applicable is supported only by a healthcare scenario the authors themselves constructed to instantiate each framework phase; the finance use case mentioned in the conclusions is never described. That closed loop—framework phase definitions are used to write a narrative, and the narrative is then cited as evidence of framework effectiveness—constitutes a real but partial circularity. The paper's own Section VI further acknowledges that the framework assumes internal model access and may not apply to black-box SaaS/API deployments, which limits but does not itself create circularity. Overall score 4 reflects one significant self-referential validation step while the framework retains independent content from its cited sources.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 1 invented entities

The central claim rests on hand-selected phases and principles, an incompletely reported literature review, and a self-authored case study. The only external grounding comes from cited standards and tools, but those are not evidence that this particular framework improves outcomes.

free parameters (2)
  • Six lifecycle phases = Not applicable (hand-selected categorical structure)
    The choice of exactly six phases (data governance, secure training, deployment, monitoring, retraining, decommissioning) is made by the authors and is not derived from a formal optimization or empirical comparison.
  • Five normative design principles = Not applicable (hand-selected categorical structure)
    Security-by-Design, Bias Auditing, Privacy Preservation, Continuous Monitoring, and Lifecycle Governance are chosen to organize the framework; no evidence is given that these five principles are necessary or sufficient.
axioms (4)
  • domain assumption NLP models have distinct lifecycle risks not covered by general AI governance frameworks.
    Motivates the entire framework; asserted in the Introduction and Related Work without a systematic comparison showing that general frameworks like NIST AI RMF or ISO 42001 fail on NLP-specific risks.
  • domain assumption The 45 selected sources from the PRISMA review are sufficient and representative.
    The review is incompletely reported: no flow diagram, search strings, or study list, so representativeness cannot be audited.
  • domain assumption The cited tools and standards work in NLP contexts as described.
    Some descriptions are not accurate to the cited works, e.g., SafeML is described as an adversarial-example detector in the case study, but the cited paper is a privacy-preserving distributed training framework.
  • ad hoc to paper A self-authored narrative case study can validate the framework's applicability.
    The healthcare scenario is constructed by the authors to fit the six phases and is presented as evidence without independent implementation or measurement.
invented entities (1)
  • SC-NLP-LMF framework no independent evidence
    purpose: Six-phase lifecycle governance model for NLP systems, intended to operationalize security, bias, privacy, and compliance practices.
    The framework is the paper's new construct. No independent validation is provided; the only illustration is a self-authored narrative case study.

pith-pipeline@v1.3.0-alltime-deepseek · 11651 in / 9826 out tokens · 100695 ms · 2026-08-03T13:54:20.846384+00:00 · methodology

0 comments
read the original abstract

Natural Language Processing (NLP) systems are increasingly used in sensitive domains such as healthcare, finance, and government, where they handle large volumes of personal and regulated data. However, these systems introduce distinct risks related to security, privacy, and regulatory compliance that are not fully addressed by existing AI governance frameworks. This paper introduces the Secure and Compliant NLP Lifecycle Management Framework (SC-NLP-LMF), a comprehensive six-phase model designed to ensure the secure operation of NLP systems from development to retirement. The framework, developed through a systematic PRISMA-based review of 45 peer-reviewed and regulatory sources, aligns with leading standards, including NIST AI RMF, ISO/IEC 42001:2023, the EU AI Act, and MITRE ATLAS. It integrates established methods for bias detection, privacy protection (differential privacy, federated learning), secure deployment, explainability, and secure model decommissioning. A healthcare case study illustrates how SC-NLP-LMF detects emerging terminology drift (e.g., COVID-related language) and guides compliant model updates. The framework offers organizations a practical, lifecycle-wide structure for developing, deploying, and maintaining secure and accountable NLP systems in high-risk environments.

Figures

Figures reproduced from arXiv: 2512.22060 by John Hastings, Sunil Arora.

Figure 1
Figure 1. Figure 1: Six Phases of Secure and Compliant NLP Lifecycle Management [PITH_FULL_IMAGE:figures/full_fig_p004_1.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

49 extracted references · 3 canonical work pages

  1. [1]

    Criminal courts’ artificial intelligence: the way it rein- forces bias and discrimination,

    M. A. Malek, “Criminal courts’ artificial intelligence: the way it rein- forces bias and discrimination,”AI and Ethics, vol. 2, no. 1, pp. 233–245,

  2. [2]

    Regulation (EU) 2024/1689 Artificial Intelligence Act,

    European Commission, “Regulation (EU) 2024/1689 Artificial Intelligence Act,” 2024. Available: https://eur-lex.europa.eu/eli/reg/ 2024/1689/oj (Accessed 2025-04-26)

    2024

  3. [3]

    Demystifying the Draft EU Artificial Intelligence Act,

    M. Veale and F. Z. Borgesius, “Demystifying the Draft EU Artificial Intelligence Act,”Computer Law Review International, vol. 22, no. 4, pp. 97–112, 2021. doi:10.9785/cri-2021-220402

    doi:10.9785/cri-2021-220402 2021

  4. [4]

    Security Update for Amazon Q Developer Extension for Visual Studio Code (Version 1.84),

    Amazon Web Services, “Security Update for Amazon Q Developer Extension for Visual Studio Code (Version 1.84),” 2025. Available: https: //aws.amazon.com/security/security-bulletins/AWS-2025-015 (Accessed 2025-07-25)

    2025

  5. [5]

    When AI Coding Assistants Turn Malicious: The Ama- zon Q Security Incident,

    C. McNamara, “When AI Coding Assistants Turn Malicious: The Ama- zon Q Security Incident,” 2025. Available: https://www.colinmcnamara. com/blog/amazon-q-ai-security-incident-analysis (Accessed 2025-08- 02)

    2025

  6. [6]

    Automation, Algorithms, and Politics— Talking to Bots: Symbiotic Agency and the Case of Tay,

    G. Neff and P. Nagy, “Automation, Algorithms, and Politics— Talking to Bots: Symbiotic Agency and the Case of Tay,”International Journal of Communication, vol. 10, no. 0, p. 17, 2016. Available: https://ijoc.org/index.php/ijoc/article/view/6277 (Accessed 2025-10-02)

    2016

  7. [7]

    Microsoft’s new AI-powered bot Tay answers your tweets and chats on GroupMe and Kik,

    S. Perez, “Microsoft’s new AI-powered bot Tay answers your tweets and chats on GroupMe and Kik,” 2016. Avail- able: https://techcrunch.com/2016/03/23/microsofts-new-ai-powered- bot-tay-answers-your-tweets-and-chats-on-groupme-and-kik/ (Ac- cessed 2025-10-02)

    2016

  8. [8]

    Artificial intelligence risk management framework (AI RMF 1.0),

    National Institute of Standards and Technology (NIST), “Artificial intelligence risk management framework (AI RMF 1.0),” Tech. Rep., 2023, doi:10.6028/NIST.AI.100-1]

    doi:10.6028/nist.ai.100-1 2023

  9. [9]

    ISO/IEC TR 24028:2020 Information Technology – Artificial Intelligence – Overview of Trustworthiness in AI,

    ISO, “ISO/IEC TR 24028:2020 Information Technology – Artificial Intelligence – Overview of Trustworthiness in AI,” International Organization for Standardization, Tech. report, 2020. Available: https://www.iso.org/standard/77608.html (Accessed 2025-06-24)

    2020

  10. [10]

    Multilayer framework for good cybersecurity practices for AI,

    N. Polemi and I. Prac ¸a, “Multilayer framework for good cybersecurity practices for AI,” The European Union Agency for Cybersecurity, Tech. Rep., 2023. Avail- able: https://www.enisa.europa.eu/publications/multilayer-framework- for-good-cybersecurity-practices-for-ai (Accessed 2025-06-24)

    2023

  11. [11]

    AI Cyber Security Code of Practice,

    Department for Science, Innovation and Technology and National Cyber Security Centre, “AI Cyber Security Code of Practice,” January

  12. [12]

    Proposal for a regulation laying down harmonised rules on artificial intelligence (Artificial Intelligence Act 2021),

    European Commission, “Proposal for a regulation laying down harmonised rules on artificial intelligence (Artificial Intelligence Act 2021),” 2021. Available: https://eur-lex.europa.eu/legal-content/EN/ TXT/?uri=CELEX%3A52021PC0206 (Accessed 2025-06-12)

    2021

  13. [13]

    ISO/IEC 42001:2023 – Information technology — Artificial intelligence — Management system,

    ISO, “ISO/IEC 42001:2023 – Information technology — Artificial intelligence — Management system,” International Organization for Standardization, Tech. report, 2023. Available: https://www.iso.org/ standard/42001 (Accessed 2025-06-22)

    2023

  14. [14]

    TFX: A TensorFlow-based production-scale machine learning platform,

    D. Bayloret al., “TFX: A TensorFlow-based production-scale machine learning platform,” in23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, ser. KDD ’17. ACM, 2017, p. 1387–1395. doi:10.1145/3097983.3098021

    arXiv 2017

  15. [15]

    Software engineering for machine learning: A case study,

    S. Amershiet al., “Software engineering for machine learning: A case study,” in2019 IEEE/ACM 41st International Conference on Software Engineering: Software Engineering in Practice (ICSE-SEIP), 2019, pp. 291–300. doi:10.1109/ICSE-SEIP.2019.00042

    arXiv 2019

  16. [16]

    MITRE ATLAS: Adversarial Threat Landscape for Artificial-Intelligence Systems,

    MITRE Corporation, “MITRE ATLAS: Adversarial Threat Landscape for Artificial-Intelligence Systems,” 2023. Available: https://atlas.mitre. org/ (Accessed 2025-06-24)

    2023

  17. [17]

    Closing the AI accountability gap: defining an end-to- end framework for internal algorithmic auditing,

    I. D. Rajiet al., “Closing the AI accountability gap: defining an end-to- end framework for internal algorithmic auditing,” in2020 Conference on Fairness, Accountability, and Transparency (FAT* ’20). ACM, 2020, p. 33–44. Available: https://doi.org/10.1145/3351095.3372873. doi:10.1145/3351095.3372873

    arXiv 2020

  18. [18]

    Fairlearn: A toolkit for assessing and improving fairness in AI,

    S. Birdet al., “Fairlearn: A toolkit for assessing and improving fairness in AI,” Microsoft, Tech. Rep. MSR-TR-2020-32, May

    2020

  19. [19]

    Frontier AI Framework (Version 1.1),

    Meta, “Frontier AI Framework (Version 1.1),” 2025. Available: https://ai. meta.com/static-resource/meta-frontier-ai-framework (Accessed 2025- 06-24)

    2025

  20. [20]

    Code of Practice for the Cyber Security of AI,

    Department for Science, Innovation and Technology, “Code of Practice for the Cyber Security of AI,” January 2025. Avail- able: https://www.gov.uk/government/publications/ai-cyber-security- code-of-practice/code-of-practice-for-the-cyber-security-of-ai (Ac- cessed 2025-10-22)

    2025

  21. [21]

    Preferred Re- porting Items for Systematic Reviews and Meta-Analyses: The PRISMA Statement,

    D. Moher, A. Liberati, J. Tetzlaff, and D. G. Altman, “Preferred Re- porting Items for Systematic Reviews and Meta-Analyses: The PRISMA Statement,”Journal of Clinical Epidemiology, vol. 62, no. 10, pp. 1006– 1012, 2009. doi:10.1016/j.jclinepi.2009.06.005

    work page doi:10.1016/j.jclinepi.2009.06.005 2009

  22. [22]

    Model cards for model reporting,

    M. Mitchellet al., “Model cards for model reporting,” inConference on Fairness, Accountability, and Transparency (FAT* ’19). ACM, 2019, p. 220–229. doi:10.1145/3287560.3287596

    arXiv 2019

  23. [23]

    Data statements for natural language processing: Toward mitigating system bias and enabling better science,

    E. M. Bender and B. Friedman, “Data statements for natural language processing: Toward mitigating system bias and enabling better science,” Transactions of the Association for Computational Linguistics, vol. 6, pp. 587–604, 12 2018. doi:10.1162/tacl a 00041

    doi:10.1162/tacl 2018

  24. [24]

    Differ- ential Privacy in Deep Learning: An Overview,

    T. Ha, T. K. Dang, T. T. Dang, T. A. Truong, and M. T. Nguyen, “Differ- ential Privacy in Deep Learning: An Overview,” in2019 International Conference on Advanced Computing and Applications (ACOMP), 2019, pp. 97–102. doi:10.1109/ACOMP.2019.00022

    arXiv 2019

  25. [25]

    Deep Learning with Differential Privacy,

    M. Abadiet al., “Deep Learning with Differential Privacy,” inPro- ceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security, ser. CCS ’16. New York, NY , USA: ACM, 2016, p. 308–318. doi:10.1145/2976749.2978318

    arXiv 2016

  26. [26]

    Fairness indicators: Scal- able infrastructure for Fair ML systems,

    C. Xu and T. Doshi, “Fairness indicators: Scal- able infrastructure for Fair ML systems,” December 11 2019. Available: https://research.google/blog/fairness-indicators- scalable-infrastructure-for-fair-ml-systems/ (Accessed 2025-05-14)

    2019

  27. [27]

    SHAP- Based Explanation Methods: A Review for NLP Interpretability,

    E. Mosca, F. Szigeti, S. Tragianni, D. Gallagher, and G. Groh, “SHAP- Based Explanation Methods: A Review for NLP Interpretability,” in Proceedings of the 29th International Conference on Computational Linguistics. International Committee on Computational Linguistics, Oct. 2022, pp. 4593–4603. Available: https://aclanthology.org/2022. coling-1.406/ (Acces...

    2022

  28. [28]

    Explainable AI for NLP: Enhancing Transparency in Sentiment Analysis and Named Entity Recognition,

    F. Elkharrat, M. Ghoniem, A. Sherif, and C. Sabty, “Explainable AI for NLP: Enhancing Transparency in Sentiment Analysis and Named Entity Recognition,” inNatural Language Processing and Information Systems, R. Ichise, Ed. Cham: Springer Nature Switzerland, 2026, pp. 116–127. doi:10.1007/978-3-031-97144-0 11

    work page doi:10.1007/978-3-031-97144-0 2026

  29. [29]

    Explainable AI in NLP: Interpretable Models for Transparent Decision Making,

    G. Shrivastava, V . Veeraiah, S. Praveenkumar, P. Pathak, T. Sharma, and A. Gupta, “Explainable AI in NLP: Interpretable Models for Transparent Decision Making,” in2024 15th International Conference on Computing Communication and Networking Technologies (ICCCNT), 2024, pp. 1–6. doi:10.1109/ICCCNT61001.2024.10726022

    Pith/arXiv arXiv 2024

  30. [30]

    SafeML: A Privacy-Preserving Byzantine-Robust Framework for Distributed Machine Learning Train- ing,

    M. Mirabi, R. K. Nikiel, and C. Binnig, “SafeML: A Privacy-Preserving Byzantine-Robust Framework for Distributed Machine Learning Train- ing,” in2023 IEEE International Conference on Data Mining Workshops (ICDMW), 2023, pp. 207–216. doi:10.1109/ICDMW60847.2023.00033

    arXiv 2023

  31. [31]

    Google’s Secure AI Framework (SAIF),

    Google, “Google’s Secure AI Framework (SAIF),” 2023. Available: https://safety.google/cybersecurity-advancements/saif (Accessed 2025- 06-23)

    2023

  32. [32]

    Publications Office, 2019

    European Commission High-Level Expert Group on Artificial Intelli- gence (AI HLEG),Ethics guidelines for trustworthy AI. Publications Office, 2019

    2019

  33. [33]

    Our AI Principles,

    Google, “Our AI Principles,” 2018. Available: https://ai.google/ principles (Accessed 2025-06-23)

    2018

  34. [34]

    Microsoft Responsible AI Standard V2,

    Microsoft, “Microsoft Responsible AI Standard V2,” 2022. Available: https://cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/ microsoft/final/en-us/microsoft-brand/documents/Microsoft- Responsible-AI-Standard-General-Requirements.pdf (Accessed 2025-05-16)

    2022

  35. [35]

    Toward Trustworthy AI Development: Mechanisms for Supporting Verifiable Claims,

    M. Brundageet al., “Toward Trustworthy AI Development: Mechanisms for Supporting Verifiable Claims,” 2020, arXiv:2004.07213 [cs.CY]

    Pith/arXiv arXiv 2020

  36. [36]

    Artificial In- telligence Trust, Risk and Security Management (AI TRiSM): Frameworks, applications, challenges and future research directions,

    A. Habbal, M. K. Ali, and M. A. Abuzaraida, “Artificial In- telligence Trust, Risk and Security Management (AI TRiSM): Frameworks, applications, challenges and future research directions,” Expert Systems with Applications, vol. 240, p. 122442, 2024. doi:10.1016/j.eswa.2023.122442

    arXiv 2024

  37. [37]

    Communication-efficient learning of deep networks from decentralized data,

    B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y. Arcas, “Communication-efficient learning of deep networks from decentralized data,” in20th International Conference on Artificial Intelligence and Statistics, vol. 54. PMLR, 20–22 Apr 2017, pp. 1273–1282. Available: https://proceedings.mlr.press/v54/mcmahan17a.html (Accessed 2025- 06-14)

    2017

  38. [39]

    Technical report on the CleverHans v2.1.0 adver- sarial examples library,

    N. Papernotet al., “Technical report on the CleverHans v2.1.0 adver- sarial examples library,” 2018, arXiv:1610.00768 [cs.LG]

    Pith/arXiv arXiv 2018

  39. [40]

    FactSheets: Increasing trust in AI services through supplier’s declarations of conformity,

    M. Arnoldet al., “FactSheets: Increasing trust in AI services through supplier’s declarations of conformity,”IBM Journal of Re- search and Development, vol. 63, no. 4/5, pp. 6:1–6:13, 2019. doi:10.1147/JRD.2019.2942288

    arXiv 2019

  40. [41]

    Mind the gap: Assessing temporal generalization in neural language models,

    A. Lazaridouet al., “Mind the gap: Assessing temporal generalization in neural language models,” inAdvances in Neural Information Processing Systems, 2021, pp. 29 348– 29 363. Available: https://proceedings.neurips.cc/paper files/paper/2021/ file/f5bf0ba0a17ef18f9607774722f5698c-Paper.pdf (Accessed 2025-06- 20)

    2021

  41. [42]

    Addressing fairness, bias, and appropriate use of artificial intelligence and machine learning in global health,

    R. R. Fletcher, A. Nakeshimana, and O. Olubeko, “Addressing fairness, bias, and appropriate use of artificial intelligence and machine learning in global health,”Frontiers in artificial intelligence, vol. 3, p. 561802,

  42. [43]

    A Review of TRiSM Frameworks in Artificial Intelligence Systems: Fundamentals, Taxonomy, Use Cases, Key Challenges and Future Directions,

    P. P. Ray, “A Review of TRiSM Frameworks in Artificial Intelligence Systems: Fundamentals, Taxonomy, Use Cases, Key Challenges and Future Directions,” 2025, techRxiv preprint: http://dx.doi.org/10.36227/ techrxiv.174913612.20443736/v1

    arXiv 2025

  43. [44]

    Tackling Trust, Risk and Security in AI Models,

    A. Litan, “Tackling Trust, Risk and Security in AI Models,”

  44. [45]

    Innovative ci/cd pipeline optimization through canary and blue-green deployment,

    S. Amgothu, “Innovative ci/cd pipeline optimization through canary and blue-green deployment,”International Journal of Computer Applica- tions, vol. 186, no. 50, pp. 1–5, 2024

    2024

  45. [2020]

    Available: https://www.microsoft.com/en-us/research/publication/ fairlearn-a-toolkit-for-assessing-and-improving-fairness-in-ai/

  46. [2021]

    doi:10.3389/frai.2020.561802

    arXiv 2020

  47. [2022]

    doi:10.1007/s43681-022-00137-9

    work page doi:10.1007/s43681-022-00137-9

  48. [2024]

    Available: https://www.gartner.com/en/articles/ai-trust-and-ai-risk (Accessed 2025-07-07)

    2025

  49. [2025]

    Available: https://www.gov.uk/government/publications/ai-cyber- security-code-of-practice (Accessed 2025-10-2)

    2025