Insight-LLM: Modular Multi-View ITD

• Insight-LLM is a multi-view fusion framework that aligns heterogeneous behavioral modalities into a shared LLM semantic space for insider threat detection.
• It employs frozen, pretrained encoders with Qformer modules to project text, sentiment, sequential, and topological signals, enabling prompt-based prediction.
• The framework’s context-aware cross-view attention dynamically reweights inputs, reducing false positives and enhancing scalability on imbalanced CERT datasets.

Insight-LLM is a modular multi-view fusion framework for insider threat detection (ITD) that is designed to analyze sparse, heterogeneous user behavior by aligning multiple behavioral modalities in a shared large-language-model semantic space and then performing context-aware cross-view fusion for prompt-based prediction (Song et al., 1 Sep 2025). It is presented as the first multi-view fusion design tailored specifically for ITD, addressing three difficulties that the paper identifies as central in this domain: scalability bottlenecks from independently trained sub-models, semantic misalignment across disparate feature spaces, and view imbalance in which high-signal modalities can suppress weaker but still threat-relevant cues (Song et al., 1 Sep 2025).

1. Problem setting and design objective

Insider threat detection is characterized by extreme sparsity of malicious behavior relative to abundant benign activity, as well as by the heterogeneous structure of the available evidence. The paper describes the relevant behavioral sources as including textual logs, event sequences, graph interactions, and affective cues. Existing ITD methods are described as predominantly single-view, which leads to limited coverage and missed anomalies, while direct transfer of multi-view learning from other domains is said to create difficulties in scalability, semantic alignment, and balance across views (Song et al., 1 Sep 2025).

Within this framing, Insight-LLM targets a specific failure mode of conventional fusion systems: weak but important indicators may be present only in a minority modality, yet those indicators can be drowned out by stronger channels if fusion is static or shallow. The framework therefore combines per-view adaptation into a shared LLM-centric latent space with a cross-view attention mechanism that dynamically reweights modalities according to context. This suggests that the model is intended not merely to concatenate heterogeneous evidence, but to regularize and arbitrate it in a way that preserves weak, collusive, or distributed threat signals.

A common misconception is that multi-view learning can be applied to insider threat detection in the same manner as in vision or generic multimodal classification. The paper argues against this view directly: in ITD, the modalities are sparse, weakly correlated, and behaviorally heterogeneous, so naïve independent sub-modeling or shallow late fusion is structurally mismatched to the task (Song et al., 1 Sep 2025).

2. Multi-view representation pipeline

The architecture begins with multi-view feature extraction from user behavior logs. The paper partitions the data into four views and assigns a frozen pretrained encoder to each one (Song et al., 1 Sep 2025).

View	Signal type	Encoder
Textual	Semantics from user messages/emails	vanilla RoBERTa encoder
Sentiment	Affective clues	sentiment-tuned RoBERTa models such as tweet-roberta-base-sentiment
Sequential	Action/event sequences for temporal anomaly detection	TS2Vec
Topological	User-resource and user-user interactions	Node2Vec

A central architectural choice is that all encoders are frozen, pretrained, and unsupervised. The paper states that this yields no additional view-specific parameter overhead, resistance to overfitting, and scalability as the number of views increases (Song et al., 1 Sep 2025). In other words, the representation backbone is not retrained per modality from scratch; instead, adaptation occurs downstream through lightweight modules.

The feature matrices are denoted by

$$\mathcal{X}_{\text{text}},\ \mathcal{X}_{\text{sent}},\ \mathcal{X}_{\text{seq}},\ \mathcal{X}_{\text{topo}}.$$

These are then passed to per-view Qformer modules called “ViewAdapter,” which constitute the manifold-alignment stage. Each view is first projected and normalized: $$\mathcal{X}'^{(v)} = \mathrm{LayerNorm}\left(\mathcal{X}^{(v)}W_{in}^{(v)}\right).$$ Using learnable query tokens and intra-view cross-attention, the aligned representation is formed as

$$Z^{(v)} = \mathrm{Intra\text{-}CrossAttn}(Q^{(v)}, \mathcal{X}'^{(v)}, \mathcal{X}'^{(v)})$$

and then refined by residual and feed-forward structure: $$\tilde{Z}^{(v)} = Z^{(v)} + Q^{(v)} + \mathrm{FFN}^{(v)}(Z^{(v)} + Q^{(v)}).$$

The stated purpose of this stage is to ensure that all views project to a shared LLM semantic space, allowing threat-relevant patterns to be probed from each modality rather than privileging only dominant channels (Song et al., 1 Sep 2025). Because the adaptation is per view and query-based, the framework avoids the parameter explosion associated with training a large bespoke model for each signal source.

3. Cross-view fusion and imbalance mitigation

After intra-view alignment, Insight-LLM concatenates the aligned tokens from all modalities: $$\mathbf{Z}_{\mathrm{cat}} = [\tilde{Z}_{\text{text}} \Vert \tilde{Z}_{\text{sent}} \Vert \tilde{Z}_{\text{seq}} \Vert \tilde{Z}_{\text{topo}}].$$ A separate set of learnable fusion queries then performs inter-view cross-attention: $$\mathcal{F} = \mathrm{Inter\text{-}CrossAttn}(Q_{\mathrm{inter}}, \mathbf{Z}_{\mathrm{cat}}, \mathbf{Z}_{\mathrm{cat}}).$$

This stage is described as “Multi-View Fusion Tuning.” Its role is to reweight views dynamically based on context, amplify weak but collusive cues, and combat view imbalance and noise heterogeneity (Song et al., 1 Sep 2025). The summary specifically notes that the reweighting is data-adaptive and context-aware, with visualizations correlating the learned weights to user and threat types.

The significance of this mechanism lies in how the paper conceptualizes imbalance. The problem is not simply unequal feature dimensionality or unequal sampling frequency. Rather, the concern is the “Matthew effect,” where a high-signal view dominates optimization and suppresses weaker modalities that may carry the only visible precursor of malicious behavior. Insight-LLM’s fusion stage is therefore a selective arbitration mechanism rather than a passive pooling layer.

The ablation findings reported in the summary are consistent with this interpretation. Text and sequence views are described as most critical, but the fusion and adaptation mechanisms are said to prevent collapse to a single dominant modality and to maintain performance when views are weak or noisy (Song et al., 1 Sep 2025). This suggests that the framework’s value is not just aggregation across views, but robustness under realistic asymmetry among them.

4. LLM integration and prompt-based prediction

Insight-LLM does not use the fused representation for direct discriminative classification. Instead, it concatenates the fused representation $\mathcal{F}$ with a tokenized natural-language prompt $\mathcal{P}$ and feeds the result into a frozen pretrained LLM such as Meta-Llama-3-8B, adapted using LoRA (Song et al., 1 Sep 2025): $X_{\text{LLM}} = [\mathcal{F}; E_{\mathcal{P}}].$

The LLM computation is represented as

$$h = \mathcal{M}_{\text{LLM}}(X_{\text{LLM}}; W_0) + \Delta W \cdot X_{\text{LLM}},$$

where $$\mathcal{X}'^{(v)} = \mathrm{LayerNorm}\left(\mathcal{X}^{(v)}W_{in}^{(v)}\right).$$0 denotes the LoRA adaptation. The label decision is then cast as prompt-based generation using class verbalizers rather than as direct classification: $$\mathcal{X}'^{(v)} = \mathrm{LayerNorm}\left(\mathcal{X}^{(v)}W_{in}^{(v)}\right).$$1 Each class, benign or malicious, is mapped to a verbalizer token, and the predicted label is determined by cosine similarity between the final LLM output and class verbalizer embeddings (Song et al., 1 Sep 2025).

The implementation details reported in the summary are specific. The backbone LLM is Meta-Llama-3-8B with hidden size 4096. LoRA is configured with rank 8, alpha 16, and dropout 0.05. Each per-view Qformer uses 4 query tokens, 8 heads, and hidden size 768. Inter-view fusion uses 8 learnable queries and 8 heads. Training is conducted with batch size 16 for 30 epochs using AdamW (Song et al., 1 Sep 2025).

The paper also emphasizes prompt and verbalizer design. Ablations are said to show that semantically matched prompts and verbalizers outperform numerical or random labels, indicating that label semantics matter for alignment in the LLM embedding space (Song et al., 1 Sep 2025). This is an important detail because it locates part of the predictive performance in the interface between fused behavioral representation and the LLM’s token semantics, not solely in the fusion network itself.

The full pipeline is summarized in four stages: per-view encoding with frozen encoders, intra-view adaptation through Qformers, cross-view fusion by attention over concatenated queries, and prompt-based LLM prediction with LoRA adaptation and label verbalizers (Song et al., 1 Sep 2025).

5. Empirical evaluation on CERT datasets

The reported experiments use two public CERT datasets, CERT r4.2 and CERT r5.2, described as containing tens of millions of events and severe class imbalance, with imbalance ratios exceeding 4,000:1 in the examples given (Song et al., 1 Sep 2025). Evaluation uses Precision, Detection Rate (Recall), False Positive Rate (FPR), and F1 score.

The comparison set includes single-view baselines such as RoBERTa, TS2Vec, and Node2Vec; heuristic fusion approaches; transformer and graph-fusion systems such as LAN and LMTracker; and LLM-based baselines (Song et al., 1 Sep 2025). The summary highlights LAN as the best competing method in the quantitative tables.

Dataset	Method	Prec	DR	FPR	F1
CERT r4.2	LAN	0.9258	0.9478	0.1222	0.9492
CERT r4.2	Insight-LLM	0.9631	0.9683	0.0476	0.9712
CERT r5.2	LAN	0.9349	0.9024	0.0865	0.9156
CERT r5.2	Insight-LLM	0.9512	0.9466	0.0496	0.9594

On CERT r4.2, the reported improvements over LAN are +4.0% Precision, +2.2% Detection Rate, $$\mathcal{X}'^{(v)} = \mathrm{LayerNorm}\left(\mathcal{X}^{(v)}W_{in}^{(v)}\right).$$2 FPR, and +2.3% F1. On CERT r5.2, the improvements are +1.7% Precision, +4.9% Detection Rate, $$\mathcal{X}'^{(v)} = \mathrm{LayerNorm}\left(\mathcal{X}^{(v)}W_{in}^{(v)}\right).$$3 FPR, and +4.8% F1 (Song et al., 1 Sep 2025). The summary explicitly emphasizes the reduction in false alarms as especially critical for ITD.

These results situate Insight-LLM’s contribution in a domain where false positives are operationally costly. The framework is not presented merely as improving ranking quality on an imbalanced benchmark; it is presented as materially reducing false positive rate while simultaneously improving recall and precision. This combination is central to the paper’s claim of state-of-the-art performance (Song et al., 1 Sep 2025).

The summary also reports parameter efficiency: LoRA and modular Qformers are said to prevent parameter blowup while maintaining scalable performance as more views are added (Song et al., 1 Sep 2025). That claim is tied directly to the architectural choice of frozen encoders and lightweight adaptation, rather than to any compression of the backbone LLM itself.

6. Interpretation, significance, and related methodological position

The paper positions Insight-LLM as an LLM-enhanced fusion framework rather than a generic LLM classifier over logs. Its novelty resides in the combination of frozen unsupervised per-view encoders, per-view Qformer alignment into an LLM-centric latent manifold, inter-view cross-attention for imbalance correction, and prompt-based LoRA-adapted prediction (Song et al., 1 Sep 2025).

A second misconception that the paper implicitly addresses is that stronger modalities should simply dominate fusion because they carry more information. Insight-LLM rejects that assumption: weak views may encode rare but decisive anomalies, especially in insider threat scenarios where malicious intent is subtle, sparse, and distributed. The cross-view attention layer is explicitly intended to amplify “weak but consistent threat cues” rather than to average them away (Song et al., 1 Sep 2025).

The framework also reflects a broader shift from shallow multimodal concatenation toward semantically aligned fusion. The paper contrasts its approach with shallow concatenation or late fusion, arguing that the ViewAdapter plus inter-view attention stack enables deeper integration in a shared LLM semantic space (Song et al., 1 Sep 2025). This suggests that the LLM is used not only as a downstream classifier but as the reference geometry into which heterogeneous signals are projected and compared.

From an operational perspective, the design choices imply a particular scalability model. New views can plausibly be added by attaching another frozen encoder and corresponding ViewAdapter, without retraining a full multi-branch architecture from scratch. Because that implication is architectural rather than directly benchmarked in the reported tables, it is best read as a modularity claim supported by the framework design and the reported parameter-efficiency rationale.

7. Place within insider threat detection research

Within insider threat detection, Insight-LLM is presented as a response to the mismatch between real behavioral heterogeneity and the narrower assumptions of many prior systems (Song et al., 1 Sep 2025). Its multi-view structure explicitly integrates textual, affective, temporal, and topological information, and its LLM-centered manifold alignment is intended to reconcile representations that would otherwise remain semantically isolated.

The paper’s reported empirical pattern is notable for two reasons. First, the strongest gains are not limited to a single metric; the model improves Precision, Detection Rate, and F1 while sharply lowering FPR. Second, the ablation results indicate that the framework does not merely exploit one dominant view. Text and sequence views are described as especially important, but fusion and adaptation are reported to preserve performance under weak or noisy views and to prevent single-modality collapse (Song et al., 1 Sep 2025).

In that sense, Insight-LLM can be understood as a semantically aligned, parameter-efficient, prompt-based ITD architecture whose principal research contribution is to turn heterogeneous behavioral evidence into a shared representational substrate for threat-sensitive reasoning. The paper’s conclusions are correspondingly specific: state-of-the-art log-based insider threat detection is achieved through LLM-based semantic alignment, cross-view attention over heterogeneous behavioral signals, frozen per-view encoders and adapters for efficient scaling, and substantial reductions in false positives on highly imbalanced real-world data (Song et al., 1 Sep 2025).