Pith. sign in

REVIEW 4 major objections 9 minor 50 references

Insight-LLM claims that fusing four behavioral views into an LLM-based pipeline detects insider threats with F1 0.9712/0.9594 and false-positive rates below 5% on the two benchmark datasets.

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 ↗

Insight-LLM fuses four behavioral views via Qformer adapters and a LoRA-tuned LLM, reporting state-of-the-art insider threat detection on CERT r4.2/r5.2, with unresolved reporting inconsistencies.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection The architecture is a sensible multi-view fusion for ITD, but the headline F1s don't reconcile with the reported precision/recall and the split leaks user identity, so the SOTA claim isn't established from the paper alone. the 4 major comments →

arxiv 2509.01509 v1 pith:MVOB2XNI submitted 2025-09-01 cs.CR

Insight-LLM: LLM-enhanced Multi-view Fusion in Insider Threat Detection

Chengyu Song , Jianming Zheng This is my paper
classification cs.CR
keywords insider threat detectionmulti-view fusionlarge language modelsQformercross-view attentionLoRAbehavioral anomaly detectionCERT datasets
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

Insider threat detection needs to catch rare, scattered signals across many kinds of user behavior. The paper argues that single-view detectors miss too much, while existing multi-view methods become expensive, semantically misaligned, and biased toward dominant views. Insight-LLM is proposed as the first multi-view fusion framework built for this problem: frozen pretrained encoders extract text, sentiment, sequence, and topology embeddings; dedicated Qformer-style adapters align them into a large language model's semantic space; cross-view attention re-weights weak threat cues; and a LoRA-tuned LLM with verbalizer tokens performs classification. On two standard insider-threat datasets, the paper reports state-of-the-art F1 and large reductions in false positives. If these results hold, the architecture offers a scalable template for fusing heterogeneous security logs without retraining large per-view models.

Core claim

Insight-LLM claims that insider threat detection can be cast as multi-view fusion and solved with a modular architecture built around a frozen large language model. Four behavioral views—text, sentiment, sequence, and topology—are each encoded by a separate frozen pretrained encoder, then projected by a dedicated ViewAdapter (a Qformer-style cross-attention module with learnable query tokens) into the LLM's shared semantic manifold. An inter-view cross-attention layer re-weights the aligned views so weak but consistent threat cues are amplified and dominant or noisy views are down-weighted. The fused token sequence, concatenated with a natural-language prompt, goes to a LoRA-tuned LLM, and t

What carries the argument

The central machinery is a four-stage pipeline: (1) frozen, unsupervised pretrained encoders produce view-specific embeddings; (2) per-view ViewAdapters—Qformer-style cross-attention modules with learnable query tokens—project each view into the LLM's semantic manifold; (3) an inter-view cross-attention layer with learnable fusion queries re-weights and combines the aligned views; (4) the fused tokens plus a natural-language prompt are fed through a LoRA-tuned LLM, and classification is read out as cosine similarity between the final hidden state and verbalizer tokens. This design lets the model leverage the LLM's pretrained semantic priors while keeping trainable parameters small.

Load-bearing premise

The load-bearing premise is that the evaluation protocol is fair and the reported numbers are accurate: the same employees appear in the training and test periods, every result is a single run with no seeds or error bars, and the F1/FPR values in different sections do not reconcile; if any of these fail, the claimed advantage over baselines is not established.

What would settle it

Re-run the exact pipeline under a user-disjoint split—train on a set of employees whose activity appears only in 2010, test on different employees in 2011—with multiple random seeds and report F1 and FPR. If F1 falls to the LAN baseline level or FPR climbs above 5%, the central claim collapses. Also reconcile Table 2's F1 0.9712/0.9594 with the parameter-study text reporting F1 0.9394 on what appears to be the same configuration.

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

If this is right

  • Insider threat detection can be built as a modular pipeline: adding a new behavioral view requires only a frozen encoder and a small adapter, avoiding the parameter explosion of training separate per-view sub-models.
  • Semantic alignment into a shared LLM manifold allows weak views like sentiment and topology to contribute without being drowned out, because cross-view attention reweights them by contextual relevance.
  • Verbalizer-based readout means class descriptions outperform arbitrary or numeric labels, so the LLM's language priors directly improve detection under label sparsity.
  • If the reported numbers hold, deployment would see roughly 42–61% fewer false alarms than the strongest compared baseline at comparable or better detection rates, reducing alert fatigue in security operations.
  • Because the text and sequence views carry most of the signal, the framework's robustness depends on keeping those encoders high-quality while using the attention modules mainly to suppress noise from weaker views.

Where Pith is reading between the lines

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

  • The training and test periods overlap in users: the same employees appear in the 2010 training data and the 2011 test data, so part of the reported gain may come from user-level memorization rather than generic anomaly detection; a user-disjoint split would quantify this.
  • The paper reports F1 0.9712 on CERT r4.2 in Table 2, while the parameter and prompt studies in Sections 5.4–5.5 report a best F1 of 0.9394 on what appears to be the same configuration; these numbers cannot both describe the same evaluation unless different splits or validation sets were used.
  • The ablation shows removing text collapses F1 to roughly 0.68 and removing sequence drops it to about 0.78, while removing topology or sentiment costs only a few points; the 'multi-view' advantage may therefore be primarily a text-plus-sequence fusion with modest complementary gains from the other views.
  • If validated on user-disjoint splits, the frozen-encoder-plus-adapter design could transfer to other sparse, heterogeneous log domains—cloud access audits, medical record access, or financial transaction logs—where labeled anomalies are rare and semantic alignment is needed to surface weak cross-modal cues.
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

4 major / 9 minor

Summary. The paper proposes Insight-LLM, a multi-view fusion framework for insider threat detection. It defines four behavioral views (textual, sentiment, sequential, topological), encodes each with a frozen pretrained encoder (RoBERTa, sentiment RoBERTa, TS2Vec, Node2Vec), maps the features into an LLM's semantic space with per-view Qformer adapters, fuses them via cross-view attention, and classifies by cosine similarity between the last hidden state and verbalizer tokens after LoRA tuning of Llama-3-8B. Experiments are conducted on CERT r4.2 and r5.2 against six baselines; the paper reports state-of-the-art F1 of 0.9712/0.9594 and FPR of 0.0476/0.0496, along with ablations, hyperparameter sensitivity, and prompt/label-word studies.

Significance. If the reported results hold, Insight-LLM is a technically plausible and timely contribution: it combines frozen per-view encoders, Qformer-based alignment into an LLM semantic space, cross-view attention, and LoRA-tuned LLM classification in a modular way that avoids training separate large models per view. The ablation study (Table 3) is informative, and the direction of the component contributions aligns with the paper's narrative. The algorithm box and sensitivity analyses are also helpful. However, the paper currently lacks the supporting evidence needed to certify the central SOTA claim: the reported F1 values are internally inconsistent, the evaluation protocol may permit user-identity leakage, and the claimed low latency/parameter overhead is not measured. If these issues are resolved (e.g., with corrected metrics, a user-disjoint split, and efficiency measurements), the contribution would be significant for the insider-threat-detection community.

major comments (4)
  1. [Table 2, §5.1] The reported F1 values do not match the harmonic mean of the same row's Precision and DR under the standard F1 definition. For r4.2, Insight-LLM: 2*0.9631*0.9683/(0.9631+0.9683)=0.9657, but Table 2 lists 0.9712. For LAN: 0.9367 vs 0.9492; for r5.2 Insight-LLM: 0.9489 vs 0.9594. The discrepancy appears in both directions in Table 3 as well (e.g., w/o Text r4.2: computed 0.6842 vs listed 0.6755). No metric definition is given anywhere. Since F1 is the headline SOTA metric in the Abstract and §5.1, the table is internally inconsistent and must be reconciled by stating the exact aggregation (per-user/session macro, sample-weighted, etc.) or correcting the numbers. This is load-bearing for the central claim.
  2. [§4.3 Datasets and Evaluation Protocol] The train/test split uses all 2010 data for training/validation and Jan-Jun 2011 for testing, with the same employees in both periods. Because CERT's threat scenarios unfold over months and inject abnormal activities per user, anomalous activities of the same malicious employees can appear in both training and test sets. The paper does not rule out this user-identity leakage, and no user-disjoint or label-disjoint evaluation is reported. The reported improvements over LAN may therefore reflect memorization of user scenarios rather than multi-view generalization. Please clarify whether any malicious user's abnormal activities appear in the training period and, if so, provide additional experiments with a user-disjoint split or equivalent safeguards. Also, all results are single-run with no seeds/error bars, so the significance of the FPR drops cannot be assessed.
  3. [Abstract; §5.4 Parameter efficiency] The abstract claims 'low latency and parameter overhead', but the paper reports no latency, throughput, memory, or parameter counts. §5.4 is titled 'Parameter efficiency' but contains only F1/FPR sensitivity sweeps over fuse-token count, LoRA rank, and α; no efficiency metric is measured. Either provide actual measurements (e.g., inference time, trainable parameters) or remove/qualify the claim.
  4. [§5.4 and §5.5] The sensitivity and prompt/label-word experiments report F1 values (e.g., 0.9394) that are substantially lower than the full-model F1 of 0.9712/0.9594 in Table 2, and the dataset/split on which these numbers are computed is not stated. Are these on a validation subset? On CERT r4.2 only? Without this context, RQ4/RQ5 results cannot be related to the main evaluation, and the final model configuration is not clearly the configuration used in §5.4 (e.g., α=0.7 vs the LoRA α=16 of §4.4). Please specify the evaluation protocol for each analysis.
minor comments (9)
  1. [Algorithm 1 and §3.4] The prose refers to 'line 7', 'line 10', 'lines 12–13', etc., but the displayed algorithm has different line numbering (e.g., feature extraction is line 2, not line 7). Please synchronize the references.
  2. [§5.5] The title 'Impact of Base LLMs' is misleading: the experiments vary prompt and verbalizer designs, not the base LLM. Rename the section accordingly.
  3. [Figure 4 caption] The caption mentions 'Parse-LLM' rather than 'Insight-LLM'.
  4. [§4.4] References [47] and [48] are for OpenLLaMA/LLaMA, not for Meta-Llama-3-8B as cited. Add the correct Llama 3 reference.
  5. [Index Terms] The index terms list 'Chain of Thought' and 'Multi-Agent', but the paper does not use CoT or multi-agent reasoning. Adjust the keywords to match content.
  6. [Eq. (16)] The LoRA update is written as ∆W·XLLM on the whole input embedding. Clarify that LoRA adapts specific attention/MLP projection matrices and align the notation with Algorithm 1.
  7. [§5.3] The authors note that the weight-allocation analysis 'requires further validation via per-sample analysis and ablation.' This caveat should be acknowledged in the conclusions or addressed with per-sample evidence.
  8. [§1 Pilot Experiments] The pilot experiment behind Figure 1 lacks dataset/split details; add a sentence describing what data produced Figure 1.
  9. [Table 2] The 'Improvements' row is not a baseline; consider moving the relative gains into the text or removing this row.

Circularity Check

0 steps flagged

No significant circularity: the SOTA claim rests on empirical benchmark comparison, not on a fitted parameter or self-citation chain.

full rationale

Insight-LLM is an empirical architecture paper. The reported F1/Precision/DR/FPR numbers come from training a concrete model (frozen text/sentiment/sequence/topology encoders, per-view Qformers, cross-view attention, LoRA-tuned Llama-3-8B with verbalizer scoring) on the 2010 CERT split and evaluating on Jan-Jun 2011. No component of the architecture is defined in terms of the target metric, no fitted value is renamed as a prediction, and no uniqueness theorem from prior work is invoked. The only self-citation ([45], Audit-LLM) appears as one item in a list of prior work using the same public datasets; it is not load-bearing for any architectural or performance claim. The information-bottleneck discussion in Sec. 3.4 is a post-hoc gloss on the cross-attention module, not the objective being optimized (cross-entropy), so no result is forced by construction. There are legitimate correctness concerns that are not circularity: Table 2's F1 values do not reconcile with the reported Precision/DR under the standard harmonic-mean definition (e.g., r4.2: 2*0.9631*0.9683/(0.9631+0.9683)=0.9657 vs reported 0.9712), Sec. 5.4 provides no latency or parameter-count measurements despite the abstract's claim, and the train/test split likely has overlapping users; these undermine reproducibility and the strength of the SOTA claim, but they do not make the derivation equivalent to its inputs. The ablation studies and external baselines give the central claim independent empirical content. No circular step is identifiable.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The framework is a heuristic architecture validated empirically and carries no invented entities. The main free parameters are the adapter and fusion token counts, LoRA rank, residual fusion weight, and the prompt/verbalizer choices, all selected on validation data, plus standard training hyperparameters that are not swept. The axioms are dominated by domain assumptions about the CERT benchmarks (reliable labels, leakage-free temporal split, pre-trained status of encoders fit on the target data) and one ad-hoc framing (the information-bottleneck justification in Section 3.4, which is not the optimized objective).

free parameters (6)
  • per-view query token count M = 4
    Number of Qformer query tokens per view; set by hand; Section 5.4 indicates token count strongly affects F1 (0.4613 at 2 tokens, 0.6150 at 6).
  • inter-view fusion token count Nq = 8
    Number of fusion queries; Section 5.4 reports a marked gain at eight tokens (F1 0.9394), i.e., selected on validation metrics.
  • LoRA rank r = 8
    Chosen from sweep over {4,6,8,16,32,64} as the best F1 in Section 5.4; smaller ranks underperform, larger ranks overfit.
  • adapter residual fusion weight alpha = 0.7
    Swept over {0.1,0.3,0.5,0.7,0.9} in Section 5.4; best at 0.7, degrading at 0.9. Distinct from the LoRA alpha equal to 16.
  • prompt and verbalizer words = "Classify user behavior types:" with class-description verbalizers
    Selected from the Section 5.5 comparison of four prompts and four label-word schemes; class-description verbalizers and task instruction prompts win.
  • class weighting, learning-rate schedule, epoch count = inverse-frequency weights; lr 5e-5, StepLR(2, 0.5), 30 epochs
    Set in Section 4.4 without a sensitivity study; these choices directly affect the imbalance handling the paper emphasizes.
axioms (5)
  • domain assumption CERT r4.2/r5.2 ground-truth labels are reliable and activity-level labels map cleanly to employee-level insider scenarios.
    Section 4.3 treats injected abnormal activities as positive labels without discussing label noise or scenario-level dependencies.
  • domain assumption The 2010 training / Jan-Jun 2011 test split is leakage-free despite overlapping users and periodic injected scenarios.
    Section 4.3 defines the split; the same employees appear in both periods, so user-identity cues could inflate results.
  • ad hoc to paper Qformer projection creates a shared LLM semantic manifold in which graph and time-series embeddings carry text-like semantics.
    Section 3.3 claims CLIP-style semantic alignment; no similarity or visualization analysis validates that heterogeneous embeddings actually cluster semantically.
  • ad hoc to paper The cross-view fusion stage approximately minimizes the information-bottleneck objective in Eq. 14.
    Section 3.4 invokes Eq. 14, but the training loss is cross-entropy with class weights; the IB Lagrangian never appears in the objective.
  • domain assumption Node2Vec and TS2Vec fit on the target data count as frozen, pre-trained encoders.
    Section 3.2 calls all encoders pre-trained and frozen; Node2Vec and TS2Vec are fit on the same CERT graphs and sequences, not transferred from other data.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Insight-LLM: LLM-enhanced Multi-view Fusion in Insider Threat Detection." pith.science (2026). https://pith.science/paper/MVOB2XNI

@misc{pith2026250901509,
  author       = {Pith},
  title        = {Pith review of: Insight-LLM: LLM-enhanced Multi-view Fusion in Insider Threat Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MVOB2XNI}},
  note         = {Machine review of arXiv:2509.01509}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Insider threat detection (ITD) requires analyzing sparse, heterogeneous user behavior. Existing ITD methods predominantly rely on single-view modeling, resulting in limited coverage and missed anomalies. While multi-view learning has shown promise in other domains, its direct application to ITD introduces significant challenges: scalability bottlenecks from independently trained sub-models, semantic misalignment across disparate feature spaces, and view imbalance that causes high-signal modalities to overshadow weaker ones. In this work, we present Insight-LLM, the first modular multi-view fusion framework specifically tailored for insider threat detection. Insight-LLM employs frozen, pre-nes, achieving state-of-the-art detection with low latency and parameter overhead.

Figures

Figures reproduced from arXiv: 2509.01509 by Chengyu Song, Jianming Zheng.

Figure 1
Figure 1. Figure 1: Performance of single-view analysis across differ [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The workflow of Insight-LLM comprises three stages: (1) partitioning user behavior logs into distinct views [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Heatmap of the average weights assigned to dif [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Performance of Parse-LLM under different log characteristics. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Performance comparison of different label word [PITH_FULL_IMAGE:figures/full_fig_p009_5.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

50 extracted references · 45 canonical work pages

  1. [1]

    2023 ponemon cost of insider threats global report,

    P. Institute, “2023 ponemon cost of insider threats global report,” Ponemon Institute, Traverse City, MI, USA, Tech. Rep.,

    work page 2023

  2. [2]

    2025 ponemon cost of insider threats global report,

    ——, “2025 ponemon cost of insider threats global report,” Ponemon Institute, Traverse City, MI, USA, Tech. Rep., 2025. [Online]. Available: https://ponemon.dtexsystems.com/

    work page 2025

  3. [3]

    Laaeb: A comprehensive log-text analysis based approach for insider threat detection,

    K. Fei, J. Zhou, Y . Zhou et al. , “Laaeb: A comprehensive log-text analysis based approach for insider threat detection,” Comput. Secur ., vol. 148, p. 104126, 2025

    work page 2025

  4. [4]

    Hetglm: Lateral movement detection by dis- covering anomalous links with heterogeneous graph neural network,

    X. Sun and J. Yang, “Hetglm: Lateral movement detection by dis- covering anomalous links with heterogeneous graph neural network,” in IEEE IPCCC . IEEE, 2022, pp. 404–411

    work page 2022

  5. [5]

    Deeplog: Anomaly detection and diagnosis from system logs through deep learning,

    M. Du, F. Li, G. Zheng, and V . Srikumar, “Deeplog: Anomaly detection and diagnosis from system logs through deep learning,” in ACM CCS. ACM, 2017, pp. 1285–1298

    work page 2017

  6. [6]

    Deep learning for un- supervised insider threat detection in structured cybersecurity data streams,

    A. Tuor, S. Kaplan, B. Hutchinson et al. , “Deep learning for un- supervised insider threat detection in structured cybersecurity data streams,” in AAAI, ser. AAAI Technical Report, vol. WS-17. AAAI Press, 2017

    work page 2017

  7. [7]

    A novel model for constructing insider threat behavior features - ubtbert ubtbert: Model for constructing insider threat behavior features user behavior text bert model,

    J. Wang, D. Zheng, D. Xie, and Y . Fang, “A novel model for constructing insider threat behavior features - ubtbert ubtbert: Model for constructing insider threat behavior features user behavior text bert model,” in ICBAR. ACM, 2024, pp. 663–668

    work page 2024

  8. [8]

    LAN: learning adaptive neighbors for real-time insider threat detection,

    X. Cai, Y . Wang, S. Xu et al., “LAN: learning adaptive neighbors for real-time insider threat detection,” IEEE TIFS , vol. 19, pp. 10 157– 10 172, 2024

    work page 2024

  9. [9]

    Roberta: A robustly optimized BERT pretraining approach,

    Y . Liu, M. Ott, N. Goyal et al. , “Roberta: A robustly optimized BERT pretraining approach,” CoRR, vol. abs/1907.11692, 2019. [Online]. Available: http://arxiv.org/abs/1907.11692

    Pith/arXiv arXiv 1907

  10. [10]

    Ts2vec: Towards universal represen- tation of time series,

    Z. Yue, Y . Wang, J. Duan et al., “Ts2vec: Towards universal represen- tation of time series,” in AAAI. AAAI Press, 2022, pp. 8980–8987

    work page 2022

  11. [11]

    node2vec: Scalable feature learning for networks,

    A. Grover and J. Leskovec, “node2vec: Scalable feature learning for networks,” in SIGKDD. ACM, 2016, pp. 855–864

    work page 2016

  12. [12]

    Uni-moe: Scaling unified multimodal llms with mixture of experts,

    Y . Li, S. Jiang, B. Hu et al. , “Uni-moe: Scaling unified multimodal llms with mixture of experts,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 47, no. 5, pp. 3424–3439, 2025

    work page 2025

  13. [13]

    Mixture of experts as representation learner for deep multi-view clustering,

    Y . Zhang, J. Cai, Z. Wu et al. , “Mixture of experts as representation learner for deep multi-view clustering,” in AAAI. AAAI Press, 2025, pp. 22 704–22 713

    work page 2025

  14. [14]

    BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models,

    J. Li, D. Li, S. Savarese, and S. C. H. Hoi, “BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models,” in ICML, ser. Proceedings of Machine Learning Research, vol. 202. PMLR, 2023, pp. 19 730–19 742

    work page 2023

  15. [15]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy et al., “Learning transferable visual models from natural language supervision,” in ICML, ser. Proceedings of Machine Learning Research, vol. 139. PMLR, 2021, pp. 8748– 8763

    work page 2021

  16. [16]

    Clip-it! language- guided video summarization,

    M. Narasimhan, A. Rohrbach, and T. Darrell, “Clip-it! language- guided video summarization,” in NIPS, 2021, pp. 13 988–14 000

    work page 2021

  17. [17]

    A new architecture for managing enterprise log data,

    A. Sah, “A new architecture for managing enterprise log data,” in LISA, A. L. Couch, Ed. USENIX, 2002, pp. 121–132

    work page 2002

  18. [18]

    Insider threat control: Using cen- tralized logging to detect data exfiltration near insider termination,

    M. Hanley and J. Montelibano, “Insider threat control: Using cen- tralized logging to detect data exfiltration near insider termination,” Tech. Rep. CMU/SEI-2011-TN-024, Oct 2011, accessed: 2025-Jul-

    work page 2011

  19. [19]

    HOLMES: real-time APT detection through correlation of suspicious information flows,

    S. M. Milajerdi, R. Gjomemo, B. Eshete et al., “HOLMES: real-time APT detection through correlation of suspicious information flows,” in SP. IEEE, 2019, pp. 1137–1152

    work page 2019

  20. [20]

    Available: https://doi.org/10.1184/R1/6574472.v1

    [Online]. Available: https://doi.org/10.1184/R1/6574472.v1

    work page doi:10.1184/r1/6574472.v1

  21. [21]

    Robust anomaly-based insider threat detection using graph neural network,

    J. Xiao, L. Yang, F. Zhong et al. , “Robust anomaly-based insider threat detection using graph neural network,” IEEE Trans. Netw. Serv. Manag., vol. 20, no. 3, pp. 3717–3733, 2023

    work page 2023

  22. [22]

    MITRE ATT&CK&AE; — attack.mitre.org,

    “MITRE ATT&CK&AE; — attack.mitre.org,” https://attack.mitre. org/, [Accessed 16-07-2025]

    work page 2025

  23. [23]

    Eigen co-occurrence matrix method for masquerade detection,

    M. Oka, Y . Oyama, and K. Kato, “Eigen co-occurrence matrix method for masquerade detection,” Publications of the Japan Society for Software Science and Technology , 2004

    work page 2004

  24. [24]

    Insider threat detection via hierarchical neural temporal point processes,

    S. Yuan, P. Zheng, X. Wu, and Q. Li, “Insider threat detection via hierarchical neural temporal point processes,” in IEEE BigData . IEEE, 2019, pp. 1343–1350

    work page 2019

  25. [25]

    A survey of multi-view representation learning,

    Y . Li, M. Yang, and Z. Zhang, “A survey of multi-view representation learning,” IEEE Trans. Knowl. Data Eng. , vol. 31, no. 10, pp. 1863– 1883, 2019

    work page 2019

  26. [26]

    Log2vec: A heterogeneous graph embedding based approach for detecting cyber threats within enter- prise,

    F. Liu, Y . Wen, D. Zhang et al. , “Log2vec: A heterogeneous graph embedding based approach for detecting cyber threats within enter- prise,” in CCS. ACM, 2019, pp. 1777–1794

    work page 2019

  27. [27]

    Multi-modal auto-encoders as joint estimators for robotics scene understanding,

    C. Cadena, A. R. Dick, and I. D. Reid, “Multi-modal auto-encoders as joint estimators for robotics scene understanding,” in Robotics: Science and Systems , 2016

    work page 2016

  28. [28]

    Multimodal deep learning,

    J. Ngiam, A. Khosla, M. Kim et al. , “Multimodal deep learning,” in ICML. Omnipress, 2011, pp. 689–696

    work page 2011

  29. [29]

    Convolutional two- stream network fusion for video action recognition,

    C. Feichtenhofer, A. Pinz, and A. Zisserman, “Convolutional two- stream network fusion for video action recognition,” in CVPR. IEEE Computer Society, 2016, pp. 1933–1941

    work page 2016

  30. [30]

    Multi-view convolutional neural networks for 3d shape recognition,

    H. Su, S. Maji, E. Kalogerakis, and E. G. Learned-Miller, “Multi-view convolutional neural networks for 3d shape recognition,” in ICCV. IEEE Computer Society, 2015, pp. 945–953

    work page 2015

  31. [31]

    Deep visual-semantic alignments for generating image descriptions,

    A. Karpathy and L. Fei-Fei, “Deep visual-semantic alignments for generating image descriptions,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 39, no. 4, pp. 664–676, 2017

    work page 2017

  32. [32]

    Learning phrase representations using RNN encoder-decoder for statistical machine translation,

    K. Cho, B. van Merrienboer, C ¸ . G ¨ulc ¸ehreet al. , “Learning phrase representations using RNN encoder-decoder for statistical machine translation,” in EMNLP. ACL, 2014, pp. 1724–1734

    work page 2014

  33. [33]

    Sequence to sequence - video to text,

    S. Venugopalan, M. Rohrbach, J. Donahue et al. , “Sequence to sequence - video to text,” in ICCV. IEEE Computer Society, 2015, pp. 4534–4542

    work page 2015

  34. [34]

    Learning a recurrent visual representation for image caption generation,

    X. Chen and C. L. Zitnick, “Learning a recurrent visual representation for image caption generation,” CoRR, vol. abs/1411.5654, 2014. [Online]. Available: http://arxiv.org/abs/1411.5654

    Pith/arXiv arXiv 2014

  35. [35]

    ITDBERT: temporal-semantic representation for insider threat detection,

    W. Huang, H. Zhu, C. Li et al. , “ITDBERT: temporal-semantic representation for insider threat detection,” in ISCC. IEEE, 2021, pp. 1–7

    work page 2021

  36. [36]

    TweetNLP: Cutting- edge natural language processing for social media,

    J. Camacho-collados, K. Rezaee, T. Riahi et al., “TweetNLP: Cutting- edge natural language processing for social media,” in EMNLP. Abu Dhabi, UAE: Association for Computational Linguistics, Dec 2022, pp. 38–49

    work page 2022

  37. [37]

    Lora: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis et al. , “Lora: Low-rank adaptation of large language models,” in ICLR. OpenReview.net, 2022

    work page 2022

  38. [38]

    Theory and application of the information bottleneck method,

    J. Lewandowsky and G. Bauch, “Theory and application of the information bottleneck method,” Entropy, vol. 26, no. 3, p. 187, 2024

    work page 2024

  39. [39]

    Contextual feature representation for image-based insider threat classification,

    S. Duan, J. Yuan, and B. Wang, “Contextual feature representation for image-based insider threat classification,” Comput. Secur ., vol. 140, p. 103779, 2024

    work page 2024

  40. [40]

    Confront insider threat: Precise anomaly detection in behavior logs based on LLM fine-tuning,

    S. Song, Y . Zhang, and N. Gao, “Confront insider threat: Precise anomaly detection in behavior logs based on LLM fine-tuning,” in COLING. ACL, 2025, pp. 8589–8601

    work page 2025

  41. [41]

    Tweet- eval: Unified benchmark and comparative evaluation for tweet classi- fication,

    F. Barbieri, J. Camacho-Collados, L. E. Anke, and L. Neves, “Tweet- eval: Unified benchmark and comparative evaluation for tweet classi- fication,” in EMNLP, ser. Findings of ACL, vol. EMNLP 2020. ACL, 2020, pp. 1644–1650

    work page 2020

  42. [42]

    Empirical evaluation of gated recurrent neural networks on sequence modeling,

    J. Chung, C ¸ . G¨ulc ¸ehre, K. Cho, and Y . Bengio, “Empirical evaluation of gated recurrent neural networks on sequence modeling,”CoRR, vol. abs/1412.3555, 2014

    Pith/arXiv arXiv 2014

  43. [43]

    A high accuracy and adaptive anomaly detection model with dual-domain graph convolutional network for insider threat detection,

    X. Li, X. Li, J. Jia et al. , “A high accuracy and adaptive anomaly detection model with dual-domain graph convolutional network for insider threat detection,” IEEE Trans. Inf. F orensics Secur ., vol. 18, pp. 1638–1652, 2023

    work page 2023

  44. [44]

    Analyzing data granularity levels for insider threat detection using machine learning,

    D. C. Le, N. Zincir-Heywood, and M. I. Heywood, “Analyzing data granularity levels for insider threat detection using machine learning,” IEEE Trans. Netw. Serv. Manag. , vol. 17, no. 1, pp. 30–44, 2020

    work page 2020

  45. [45]

    Audit-llm: Multi-agent collaboration for log-based insider threat detection,

    C. Song, L. Ma, J. Zheng et al., “Audit-llm: Multi-agent collaboration for log-based insider threat detection,” CoRR, vol. abs/2408.08902, 2024

    Pith/arXiv arXiv 2024

  46. [46]

    Few-shot insider threat detection,

    S. Yuan, P. Zheng, X. Wu, and H. Tong, “Few-shot insider threat detection,” in CIKM. ACM, 2020, pp. 2289–2292

    work page 2020

  47. [47]

    Openllama: An open reproduction of llama,

    X. Geng and H. Liu, “Openllama: An open reproduction of llama,” May 2023. [Online]. Available: https://github.com/openlm-research/ open llama

    work page 2023

  48. [48]

    Bridging the gap: A pragmatic approach to generating insider threat data,

    J. Glasser and B. Lindauer, “Bridging the gap: A pragmatic approach to generating insider threat data,” in 2013 IEEE SP . IEEE Computer Society, 2013, pp. 98–104

    work page 2013

  49. [50]

    Llama: Open and efficient foundation language models,

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozi`ere, N. Goyal, E. Hambro, F. Azhar et al., “Llama: Open and efficient foundation language models,” arXiv preprint arXiv:2302.13971, 2023

    Pith/arXiv arXiv 2023

  50. [2023]

    Available: https://ponemonsullivanreport.com/2023/ 10/cost-of-insider-risks-global-report-2023/

    [Online]. Available: https://ponemonsullivanreport.com/2023/ 10/cost-of-insider-risks-global-report-2023/

    work page 2023

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.