Pith. sign in

REVIEW 3 major objections 48 references

Higher agreement among embedding models singles out cross-protocol signalling contradictions, shrinking the set that needs inspection from 44% of records to under 1%.

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 ↗

Multi-embedding consensus on fused SS7/Diameter/GTP records ranks synthetic cross-protocol inconsistencies far above original traffic, reducing the inspection set from 44% to under 1% at full agreement.

T0 review reviewed 2026-07-14 challenge →

load-bearing objection Solid, carefully scoped triage method for a real telecom pain point; the ranking works cleanly on their synthetics, but the operational claim still hangs on whether field-swap outliers equal real cross-protocol inconsistencies. the 3 major comments →

arxiv 2603.15344 v2 pith:WKQ3NSGF submitted 2026-03-16 cs.CR

Unsupervised Cross-Protocol Anomaly Analysis in Mobile Core Networks via Multi-Embedding Models Consensus

Aayush Garg , Orlando Amaral Cejas This is my paper
classification cs.CR
keywords cross-protocol anomaly detectionmobile core networksSS7DiameterGTPmulti-embedding consensusunsupervised detectionsynthetic field swaps
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.

The reading

Mobile cores run SS7, Diameter and GTP side by side, so many security problems only appear when the three are checked together for the same subscriber and minute. Labels for real multi-protocol attacks are almost never available, so the authors fuse those messages into per-minute records, turn each record into text, embed it with six different models, and let each model’s unsupervised detector vote. The number of models that vote “anomalous” becomes a consensus score. To test the score without real labels they plant synthetic contradictions by swapping one carefully chosen field group at a time; each protocol message stays valid on its own, yet the joint view becomes inconsistent. On 219 294 records any single model flags nearly half the traffic, yet full agreement of all six models flags only 0.97 percent—and every one of those records is synthetic. The result gives operators a practical, label-free way to rank the tiny fraction of windows most likely to hide the cross-protocol problems that single-protocol rules miss.

Core claim

On a corpus of 219 294 fused records, the consensus score that counts how many of six embedding models flag a record is strongly associated with synthetic cross-protocol inconsistencies: for consensus 1–4 the odds that a flagged record is synthetic are hundreds of times higher than for original traffic, and for consensus ≥5 every flagged record is synthetic (p-values effectively zero). Cosine distances between synthetic and original decision vectors also rise with consensus, showing clearer separation in embedding space. The practical consequence is that the candidate set for human inspection can be reduced from 44 % of traffic to 0.97 % while retaining essentially all of the planted contrad

What carries the argument

The consensus score—the integer count of embedding models whose Isolation Forest labels a fused per-subscriber, per-minute record as anomalous—together with the single field-group swap generator that creates the thirteen families of protocol-plausible but jointly contradictory synthetic anomalies used for evaluation.

Load-bearing premise

The central ranking claim stands only if single field-group swaps between otherwise normal fused records are a good enough proxy for the cross-protocol inconsistencies that actually matter in live networks.

What would settle it

Collect a set of real, operator-confirmed cross-protocol incidents, score them with the same six-model pipeline, and check whether high consensus (k ≥ 5) recovers most of those incidents while still leaving almost no original traffic in the flagged set; if recovery is low or background leakage is high, the prioritisation claim fails.

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

If this is right

  • Operators can restrict manual review to the high-consensus slice (for example S ≥ 5, 3.9 % of the corpus) and still capture essentially every synthetic cross-protocol contradiction.
  • Existing single-protocol rule engines can stay in place; the ensemble supplies the missing joint view without any new labelled incidents.
  • The same fusion-plus-consensus pipeline can be reused on any other multi-protocol control plane once messages are serialised as text.
  • Raising the consensus threshold produces a predictable precision–volume trade-off that can be matched to available analyst capacity.

Where Pith is reading between the lines

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

  • If real multi-protocol attacks produce similar embedding-space separation, the method could surface zero-day interconnect fraud without waiting for signature updates.
  • Context-window limits already force partial consensus for longer records; larger windows or hierarchical serialisation could raise the ceiling of full agreement.
  • The same label-free ranking idea may apply to any domain that must check multiple related event streams for mutual consistency (for example multi-log system forensics).
  • Extending the snapshot mutations to sequences of fused minutes could capture multi-step attacks that the current single-window design misses.
Share X Bluesky LinkedIn Reddit HN

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 / 0 minor

Summary. The paper proposes an unsupervised pipeline for cross-protocol anomaly analysis in mobile core networks. SS7, Diameter, and GTP messages are fused per subscriber and minute into JSON records, serialized as text, embedded by six models, and scored by Isolation Forest; a consensus count S equals the number of models that flag a record. Evaluation uses 13 families of single field-group swaps that keep each message valid while inducing fused-view contradictions. On 219,294 records, 44.15% are flagged by at least one model but only 0.97% reach full agreement; for k=1–4 the odds that a flagged record is synthetic are hundreds of times larger than for original records, and for k≥5 all flagged records are synthetic (Fisher p-values extremely small). Cosine distances between synthetic and original decision vectors are also reported. The authors conclude that multi-embedding consensus can prioritize a much smaller candidate set of cross-protocol inconsistencies for inspection.

Significance. Cross-protocol signalling inconsistencies are a recognized operational problem for which labeled incidents are scarce; an unsupervised ranking method that requires no proprietary schemas and no labels at fit time is therefore of practical interest. The fusion construction, the explicit list of consistency expectations (§III), the 13 mutation families (§IV-B), and the clean statistical association on held-out synthetics (Tables IV–VI) are concrete contributions. The work is carefully scoped: it does not claim the synthetics are real attacks, reports near-zero background leakage, and lists construct-validity threats in §VII. If the ranking proved useful on real traffic, the prioritization reduction (from ~44% to <1%) would be operationally valuable. The main limitation is that the central claim is currently supported only inside the synthetic regime.

major comments (3)
  1. §IV-B and Tables IV–VI: the central ranking claim rests on association between consensus S and the 13 single field-group swaps. Those swaps produce large, localized lexical differences in the serialized JSON (entire CGGT/CDGT, Session-Id, TEID, Visited-PLMN-Id, PDN IP, etc.). Isolation Forest on text embeddings can flag such records simply as rare token-level outliers relative to the background, without recovering the semantic contradictions listed in §III. The paper acknowledges construct validity in §VII, yet the reported odds ratios (236–572) and perfect separation at k≥5 are exactly what any detector sensitive to rare field values would produce. Without evidence that the same ranking recovers non-swap inconsistencies (multi-field, temporal, or real incidents), the prioritization claim for operational traffic remains unanchored. A minimal strengthening would be an ablation that (a) in
  2. §V-C and Table VII: the cosine-distance analysis is performed on the six-dimensional binary decision vectors of the models, not on the embedding vectors themselves. The abstract and §VI-E wording (“distances between synthetic and original records increase with the consensus level, suggesting clearer separation in embedding space”) therefore overstates what is measured. Decision-vector disagreement is expected once S is high by construction; it does not independently demonstrate geometric separation of the embeddings. Either recompute distances in the actual embedding spaces or revise the claim to match the decision-vector construction.
  3. §V-B and Table V: two of the six models process only ~26% of synthetic and ~12% of background records because of context-window limits; the median record therefore receives decisions from only four models. Consensus S is consequently often computed over a partial ensemble, and the maximum achievable S is bounded by coverage. The paper reports missing rates but does not show that the ranking remains stable when restricted to the four full-coverage models, nor does it normalize S by N(r). Because the operational claim is that high consensus prioritizes a small candidate set, sensitivity of the odds ratios and of the k≥5 perfect separation to the partial-coverage models should be reported.

Circularity Check

1 steps flagged

No load-bearing definitional circularity; mild evaluation self-reference only—association with planted field-swap synthetics is the evaluation target itself, not a forced first-principles prediction.

specific steps
  1. fitted input called prediction [§V-C Evaluation (training pool) and §VI-B/D (precision / odds at S≥k)]
    "For each model, we form an unlabeled training pool that contains background embeddings and a small proportion of synthetic embeddings (5% of background embeddings). Including a small proportion of synthetic embeddings in the pool lets the detector learn a density landscape where the synthetic rows are rare... We fit an unsupervised Isolation Forest on this pool with a fixed contamination rate of 0.01... For k=1−4, the odds that a flagged record is synthetic are hundreds of times larger than for original records, and for k≥5 all flagged records are synthetic"

    A small unlabeled fraction of the same synthetic family distribution used as the positive control is present in the detector’s training pool. Recovery of held-out members of those families at high consensus is therefore partly facilitated by the density landscape the detector was shown, rather than being a fully out-of-distribution prediction. The effect is limited (held-out split by family; contamination fixed; background leakage near zero) and does not collapse the consensus ranking to a definitional identity, but it is a mild fitted-input → recovery loop.

full rationale

This is an empirical unsupervised-detection paper, not a first-principles derivation. The method (per-subscriber/minute fusion → multi-model text embeddings → Isolation Forest → consensus count S) is defined independently of the evaluation labels. The central reported quantities—odds ratios 236–572 for S≥k, perfect separation at k≥5, near-zero background leakage—are empirical measurements on held-out synthetic records, not quantities algebraically identical to fitted inputs. Including a small unlabeled synthetic fraction (5%) in the Isolation Forest training pool with fixed contamination 0.01 is a mild optimistic bias risk, but the paper holds out a disjoint synthetic subset split by mutation family and does not treat fitted contamination as a predicted attack rate. Self-citations (Garg et al. on mutation evaluation and text serialization) supply methodological parallels only; they do not import uniqueness theorems or force the consensus ranking. The acknowledged construct-validity gap (single field-group swaps as proxies for operational inconsistencies, §VII) is a validity threat, not circularity of derivation. Score 2 reflects that one minor evaluation self-reference without elevating it to a forced reduction of the central claim.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 2 invented entities

The central ranking claim rests on a small set of modelling choices (contamination rate, synthetic mix fraction, Isolation Forest, six particular embedders, 13 hand-chosen field groups) and on the domain assumption that single-field swaps are a valid evaluation proxy. No new physical entities are postulated; the free parameters are standard hyper-parameters of the unsupervised pipeline.

free parameters (4)
  • Isolation Forest contamination rate = 0.01
    Fixed at 0.01; directly controls the fraction of points labelled anomalous and therefore the consensus counts.
  • synthetic mix fraction in training pool = 0.05
    5 % of background embeddings; chosen by hand to keep synthetics rare while remaining label-free.
  • choice and number of embedding models = 6 models
    Exactly the six listed Ollama models; different selections would alter coverage and consensus distribution.
  • 13 field-group mutation families = 13
    Hand-selected by the authors (and industrial partners) as security-relevant; the evaluation statistics are defined with respect to these families.
axioms (3)
  • domain assumption Single field-group swaps preserve per-protocol validity while inducing operationally meaningful cross-protocol contradictions.
    Stated throughout §IV-B and used as the sole positive-control class; construct validity rests on it (§VII).
  • domain assumption Isolation Forest on standardized embedding vectors yields a useful binary anomaly decision for each model.
    Adopted without ablation against other unsupervised detectors (§V-B).
  • ad hoc to paper Records whose tokenized length exceeds a model’s context window may be dropped without biasing the consensus ranking.
    Two of the six models process only ~26 % of synthetics; missingness is reported but not corrected for (§VI-C).
invented entities (2)
  • consensus score S(r) no independent evidence
    purpose: Integer count of embedding models that flag a fused record; used as the ranking variable.
    Defined in §IV-D; no independent existence outside this ranking scheme.
  • per-subscriber per-minute fused record no independent evidence
    purpose: Aggregation unit that makes cross-protocol checks possible.
    Constructed in §IV-A; the entire analysis is defined on these objects.

reviewed 2026-07-14 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Unsupervised Cross-Protocol Anomaly Analysis in Mobile Core Networks via Multi-Embedding Models Consensus." pith.science (2026). https://pith.science/paper/WKQ3NSGF

@misc{pith2026260315344,
  author       = {Pith},
  title        = {Pith review of: Unsupervised Cross-Protocol Anomaly Analysis in Mobile Core Networks via Multi-Embedding Models Consensus},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WKQ3NSGF}},
  note         = {Machine review of arXiv:2603.15344}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Mobile core networks rely on several signalling protocols in parallel, such as SS7, Diameter, and GTP, so many security-relevant problems become visible only when their interactions are analyzed jointly. At the same time, labeled examples of real attacks and cross-protocol misconfigurations are scarce, which complicates supervised detection. We therefore study unsupervised cross-protocol anomaly analysis on fused representations that combine SS7, Diameter, and GTP signalling. For each subscriber, we aggregate messages into per-minute fused records, serialize each record as text, embed it with several models, and apply unsupervised anomaly detection. We then assign each record a consensus score equal to the number of embedding models that flag it as anomalous. For evaluation, we generate cross-protocol-plausible synthetic anomalies by swapping one field group at a time between pairs of records, preserving per-message validity while making the fused view contradictory. On 219,294 fused records, 44.15% are flagged by at least one model, but only 0.97% reach full agreement across all six. Higher consensus is strongly associated with synthetic records, where for k=1-4 the odds that a flagged record is synthetic are hundreds of times greater than for original records, and for k>=5 all flagged records are synthetic, with extremely small p-values. Cosine distances between synthetic and original records also increase with consensus, suggesting clearer separation in embedding space. These results support the use of multi-embedding consensus to prioritize a much smaller set of candidate cross-protocol inconsistencies for further inspection.

Figures

Figures reproduced from arXiv: 2603.15344 by Aayush Garg, Orlando Amaral Cejas.

Figure 1
Figure 1. Figure 1: Anonymized signalling samples. On the left, we show individual fragments of SS7 and Diameter protocols. On the right, [PITH_FULL_IMAGE:figures/full_fig_p007_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

48 extracted references · 1 canonical work pages

  1. [3]

    An empirical study of deep learning-based ss7 attack detection,

    Y . Guo, O. Ermis, Q. Tang, H. Trang, and A. De Oliveira, “An empirical study of deep learning-based ss7 attack detection,”Information, vol. 14, no. 9, p. 509, 2023

    2023

  2. [4]

    Diameter Security: An Auditor’s Viewpoint,

    S. Mashukov and Positive Technologies, Russia, “Diameter Security: An Auditor’s Viewpoint,”Journal of ICT Standardization, vol. 5, no. 1, pp. 53–68, 2017. [Online]. Available: http://www.riverpublishers.com/ journal read html article.php?j=JICTS/5/1/3

    2017

  3. [5]

    Protocol Misuse in Mobile Networks: Real-World Case Studies and Security Risks

    “Protocol Misuse in Mobile Networks: Real-World Case Studies and Security Risks.” [Online]. Available: https://www.p1sec.com/blog/proto col-misuse-in-mobile-networks-case-studies-from-the-wild

  4. [6]

    Enea Adaptive Signalling Firewall

    “Enea Adaptive Signalling Firewall.” [Online]. Available: https: //www.enea.com/solutions/signalling-security/signalling-firewall/

  5. [7]

    Network Intrusion Detection with Limited Labeled Data,

    S. Lotfi, M. Modirrousta, S. Shashaani, S. Amini, and M. Shoore- hdeli, “Network Intrusion Detection with Limited Labeled Data,”arXiv preprint arXiv:2209.03147, 2022

    Pith/arXiv arXiv 2022

  6. [8]

    Unsupervised Machine Learning Methods for Anomaly Detection in Network Packets,

    H. Park, D. Shin, C. Park, J. Jang, and D. Shin, “Unsupervised Machine Learning Methods for Anomaly Detection in Network Packets,”Electronics, vol. 14, no. 14, p. 2779, Jan. 2025, publisher: Multidisciplinary Digital Publishing Institute. [Online]. Available: https://www.mdpi.com/2079-9292/14/14/2779

    2025

  7. [9]

    Survey of intrusion detection systems: techniques, datasets and challenges,

    A. Khraisat, I. Gondal, P. Vamplew, and J. Kamruzzaman, “Survey of intrusion detection systems: techniques, datasets and challenges,” Cybersecurity, vol. 2, no. 1, p. 20, Jul. 2019. [Online]. Available: https://doi.org/10.1186/s42400-019-0038-7

    doi:10.1186/s42400-019-0038-7 2019

  8. [10]

    Overview on Intrusion Detection Systems for Computers Networking Security,

    L. Diana, P. Dini, and D. Paolini, “Overview on Intrusion Detection Systems for Computers Networking Security,”Computers, vol. 14, no. 3, p. 87, Mar. 2025, publisher: Multidisciplinary Digital Publishing Institute. [Online]. Available: https://www.mdpi.com/2073-431X/14/3/8 7

    2025

  9. [11]

    A survey of network-based intrusion detection data sets,

    M. Ring, S. Wunderlich, D. Scheuring, D. Landes, and A. Hotho, “A survey of network-based intrusion detection data sets,”Computers & Security, vol. 86, pp. 147–167, Sep. 2019. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S016740481930118X

    2019

  10. [12]

    Unsupervised anomaly detection algorithms on real-world data: how many do we need?

    R. Bouman, Z. Bukhsh, and T. Heskes, “Unsupervised anomaly detection algorithms on real-world data: how many do we need?”J. Mach. Learn. Res., vol. 25, no. 1, pp. 105:5199–105:5232, Jan. 2024

    2024

  11. [13]

    Rule-based anomaly detection model with stateful correlation enhancing mobile network security

    R. Afzal and R. K. Murugesan, “Rule-based anomaly detection model with stateful correlation enhancing mobile network security.”Intelligent Automation & Soft Computing, vol. 31, no. 3, 2022

    2022

  12. [14]

    Learning from what we know: How to perform vulnerability prediction using noisy historical data,

    A. Garg, R. Degiovanni, M. Jimenez, M. Cordy, M. Papadakis, and Y . L. Traon, “Learning from what we know: How to perform vulnerability prediction using noisy historical data,”Empir. Softw. Eng., vol. 27, no. 7, p. 169, 2022. [Online]. Available: https://doi.org/10.1007/s10664-022-10197-4

    work page doi:10.1007/s10664-022-10197-4 2022

  13. [15]

    Learning to predict vulnerabilities from vulnerability-fixes: A machine translation approach,

    ——, “Learning to predict vulnerabilities from vulnerability-fixes: A machine translation approach,”CoRR, vol. abs/2012.11701, 2020. [Online]. Available: https://arxiv.org/abs/2012.11701

    Pith/arXiv arXiv 2012

  14. [16]

    A cnn-based semi-supervised learning approach for the detection of ss7 attacks,

    O. Ermis, C. Feltus, Q. Tang, H. Trang, A. De Oliveira, C. Nguyen, and A. Hirtzig, “A cnn-based semi-supervised learning approach for the detection of ss7 attacks,” inInternational Conference on Information Security Practice and Experience. Springer, 2022, pp. 345–363

    2022

  15. [17]

    Analysis and application of ss7 core net- work threat detection,

    Z. Xu, Y . Wang, and X. Ge, “Analysis and application of ss7 core net- work threat detection,” inThird International Conference on Computer Science and Communication Technology (ICCSCT 2022), vol. 12506. SPIE, 2022, pp. 353–360

    2022

  16. [18]

    Attacks to mobile networks using ss7 vulnerabilities: a real traffic analysis,

    L. O. H. de Carvalho Macedo and M. E. M. Campista, “Attacks to mobile networks using ss7 vulnerabilities: a real traffic analysis,” Telecommunication Systems, vol. 83, no. 3, pp. 253–265, 2023

    2023

  17. [19]

    Survey on threats and attacks on mobile networks,

    S. Mavoungou, G. Kaddoum, M. Taha, and G. Matar, “Survey on threats and attacks on mobile networks,”IEEE Access, vol. 4, pp. 4543–4572, 2016

    2016

  18. [20]

    Security penetration test framework for the diameter protocol,

    F. R. Carlson, “Security penetration test framework for the diameter protocol,”arXiv preprint arXiv:1407.6508, 2014

    Pith/arXiv arXiv 2014

  19. [21]

    Diameter security: An auditor’s viewpoint,

    S. Mashukov, “Diameter security: An auditor’s viewpoint,”Journal of ICT Standardization, vol. 5, no. 1, pp. 53–68, 2017

    2017

  20. [22]

    Device-centric detection and mitigation of diameter signaling attacks against mobile core,

    Z. Tan, B. Ding, Z. Zhang, Q. Li, Y . Guo, and S. Lu, “Device-centric detection and mitigation of diameter signaling attacks against mobile core,” in2021 IEEE Conference on Communications and Network Security (CNS). IEEE, 2021, pp. 29–37

    2021

  21. [23]

    User location tracking attacks for lte networks using the interworking functionality,

    S. Holtmanns, S. P. Rao, and I. Oliver, “User location tracking attacks for lte networks using the interworking functionality,” in2016 IFIP Networking conference (IFIP Networking) and workshops. IEEE, 2016, pp. 315–322

    2016

  22. [24]

    Real threats using gtp protocol and countermeasures on a 4g mobile grid computing environment,

    S. Park, S. Kim, K. Son, H. Kim, J. Park, and K. Yim, “Real threats using gtp protocol and countermeasures on a 4g mobile grid computing environment,”International Journal of Web and Grid Services, vol. 13, no. 1, pp. 3–24, 2017

    2017

  23. [25]

    Detecting iot botnet in 5g core network using machine learning,

    Y .-E. Kim, M.-G. Kim, and H. Kim, “Detecting iot botnet in 5g core network using machine learning,”Computers, Materials and Continua, vol. 72, no. 3, pp. 4467–4488, 2022

    2022

  24. [26]

    Machine learning based signaling ddos detection system for 5g stand alone core network,

    S. Park, B. Cho, D. Kim, and I. You, “Machine learning based signaling ddos detection system for 5g stand alone core network,”Applied Sciences, vol. 12, no. 23, p. 12456, 2022

    2022

  25. [27]

    Sdn- based gtp-u traffic analysis for 5g networks,

    D. Pineda, R. Harrilal-Parchment, K. Akkaya, and A. Perez-Pons, “Sdn- based gtp-u traffic analysis for 5g networks,” inNOMS 2023-2023 IEEE/IFIP Network Operations and Management Symposium. IEEE, 2023, pp. 1–4

    2023

  26. [28]

    On the coupling between vulnerabilities and llm-generated mutants: A study on vul4j dataset,

    A. Garg, R. Degiovanni, M. Papadakis, and Y . L. Traon, “On the coupling between vulnerabilities and llm-generated mutants: A study on vul4j dataset,” inIEEE Conference on Software Testing, Verification and Validation, ICST 2024, Toronto, ON, Canada, May 27-31, 2024. IEEE, 2024, pp. 305–316. [Online]. Available: https://doi.org/10.1109/ICST60714.2024.00035

    doi:10.1109/icst60714.2024.00035 2024

  27. [29]

    Q.713:Signalling connection control part formats and codes

    “Q.713:Signalling connection control part formats and codes.” [Online]. Available: https://www.itu.int/rec/T-REC-Q.713

  28. [30]

    Specification # 29.002

    “Specification # 29.002.” [Online]. Available: https://portal.3gpp.org/de sktopmodules/Specifications/SpecificationDetails.aspx?specificationId= 1585

  29. [31]

    “TS 123 003 - V16.3.0 - Digital cellular telecommunications sys- tem (Phase 2+) (GSM); Universal Mobile Telecommunications System (UMTS); LTE; 5G; Numbering, addressing and identification (3GPP TS 23.003 version 16.3.0 Release 16).”

  30. [32]

    Diameter Base Protocol,

    V . Fajardo, J. Arkko, J. A. Loughney, and G. Zorn, “Diameter Base Protocol,” Internet Engineering Task Force, Request for Comments RFC 6733, Oct. 2012, num Pages: 152. [Online]. Available: https://datatracker.ietf.org/doc/rfc6733

    2012

  31. [33]

    ETSI TS 129 272 V17.2.0 (2022-05) Universal Mobile Telecommunications

    “ETSI TS 129 272 V17.2.0 (2022-05) Universal Mobile Telecommunications.” [Online]. Available: https://www.intertekinfo rm.com/en-gb/standards/

    2022

  32. [34]

    Signalling Security in Telecom SS7/Diameter/5G|ENISA,

    “Signalling Security in Telecom SS7/Diameter/5G|ENISA,” Sep

  33. [35]

    Available: https://www.enisa.europa.eu/publications/sig nalling-security-in-telecom-ss7-diameter-5g

    [Online]. Available: https://www.enisa.europa.eu/publications/sig nalling-security-in-telecom-ss7-diameter-5g

  34. [36]

    Q.704:Signalling network functions and messages

    “Q.704:Signalling network functions and messages.” [Online]. Available: https://www.itu.int/rec/T-REC-Q.704

  35. [37]

    Specification # 23.002

    “Specification # 23.002.” [Online]. Available: https://portal.3gpp.org/de sktopmodules/Specifications/SpecificationDetails.aspx?specificationId= 728

  36. [38]

    Q.773:Transaction capabilities formats and encoding

    “Q.773:Transaction capabilities formats and encoding.” [Online]. Available: https://www.itu.int/rec/T-REC-Q.773

  37. [39]

    ETSI TS 129 274

    “ETSI TS 129 274.” [Online]. Available: https://store.accuristech.com/ standards/etsi-ts-129-274?product id=2208669&srsltid=AfmBOopGku 6zMszUYfV oYp4YwMP2Vakx8L6ZPzYwOFCfam5yXXaeAt4

  38. [40]

    IR.88 EPS Roaming Guidelines Version 29.0,

    “IR.88 EPS Roaming Guidelines Version 29.0,” Dec. 2021. [Online]. Available: https://www.gsma.com/newsroom/gsma resources/ir-88-eps -roaming-guidelines-version-27-1/

    2021

  39. [41]

    Communications Security, Reliability, and Interoperability Council VI|Federal Communications Commission

    “Communications Security, Reliability, and Interoperability Council VI|Federal Communications Commission.” [Online]. Available: https://www.fcc.gov/about-fcc/advisory-committees/communications-s ecurity-reliability-and-interoperability-council

  40. [42]

    ETSI TS 129 281 V17.4.0 (2022-10) Universal Mobile Telecommunications

    “ETSI TS 129 281 V17.4.0 (2022-10) Universal Mobile Telecommunications.” [Online]. Available: https://www.intertekinfo rm.com/en-ca/standards/

    2022

  41. [43]

    “TS 123 401 - V17.11.0 - LTE; General Packet Radio Service (GPRS) enhancements for Evolved Universal Terrestrial Radio Access Network (E-UTRAN) access (3GPP TS 23.401 version 17.11.0 Release 17).”

  42. [44]

    Enabling efficient assertion inference,

    A. Garg, R. Degiovanni, F. Molina, M. Cordy, N. Aguirre, M. Papadakis, and Y . L. Traon, “Enabling efficient assertion inference,” in34th IEEE International Symposium on Software Reliability Engineering, ISSRE 2023, Florence, Italy, October 9-12, 2023. IEEE, 2023, pp. 623–634. [Online]. Available: https://doi.org/10.1109/ISSRE59848.2023.00039

    doi:10.1109/issre59848.2023.00039 2023

  43. [45]

    Bge m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation,

    J. Chen, S. Xiao, P. Zhang, K. Luo, D. Lian, and Z. Liu, “Bge m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation,” 2024

    2024

  44. [46]

    Jina Embeddings 2: 8192-Token General-Purpose Text Embeddings for Long Documents,

    M. G ¨unther, J. Ong, I. Mohr, A. Abdessalem, T. Abel, M. K. Akram, S. Guzman, G. Mastrapas, S. Sturua, B. Wang, M. Werk, N. Wang, and H. Xiao, “Jina Embeddings 2: 8192-Token General-Purpose Text Embeddings for Long Documents,” Feb. 2024, arXiv:2310.19923 [cs]. [Online]. Available: http://arxiv.org/abs/2310.19923

    Pith/arXiv arXiv 2024

  45. [47]

    Introducing Nomic Embed: A Truly Open Embedding Model

    “Introducing Nomic Embed: A Truly Open Embedding Model.” [Online]. Available: https://www.nomic.ai/blog/posts/nomic-embed-tex t-v1

  46. [48]

    snowflake-arctic-embed2:568m

    “snowflake-arctic-embed2:568m.” [Online]. Available: https://ollama.c om/snowflake-arctic-embed2:568m

  47. [49]

    Snowflake launches text-embedding model for retrieval use cases

    “Snowflake launches text-embedding model for retrieval use cases.” [Online]. Available: https://www.snowflake.com/content/snowflake-site/ global/en/blog/introducing-snowflake-arctic-embed-snowflakes-state-o f-the-art-text-embedding-family-of-models

  48. [50]

    unclemusclez/jina-embeddings-v2-base-code:q8

    “unclemusclez/jina-embeddings-v2-base-code:q8.” [Online]. Available: https://ollama.com/unclemusclez/jina-embeddings-v2-base-code:q8

This paper was first reviewed by grok-4.5 on July 14, 2026.