acceptodds
Under review as a conference paper at ICLR 2027

SecBFF: Topology-Aware Security LLMs

Abstract

Large Language Models (LLMs) have emerged as powerful tools for reasoning over temporal event sequences in security applications. However, they are not designed to directly leverage the broader interaction topology among the entities involved, and therefore miss structural signals that are often highly indicative of anomalous behavior. To address this challenge, we propose SecBFF, a framework that makes LLMs topology-aware and improving predictive performance. A key feature of SecBFF is that it is effective even when topological information covers only a small fraction of the data. SecBFF uses Topology-Injected Masked Fine-Tuning, a strategy in which the LLM learns to condition its temporal reasoning on continuous topological tokens, without costly vocabulary modifications. Across three real-world tasks spanning Cyber-Security, Blockchain and Access Violation, SecBFF significantly improves existing approaches on both real-time and full-sequence classification. We further show that SecBFF is more robust to variation in event-sequence length.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Security, LLMs, Graphs

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_KiLLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposed the SecBFF framework which tries to address the temporal security event detection when graph coverage is not complete by combining graph structure with LLMs. The proposed method is topology-injected masked fine-tuning. It replaces the text embeddings of entities which are covered by pre-trained GFMs with graph embeddings with linear projection and layered normalization. For entities without graph embedding, the original text embedding will be kept. Authors evaluated the method on three datasets with five detection tasks. Authors claimed that using SecBFF, it outperformed text-only models and prefix-based graph LLMs in all tasks.

Critical strengths

Even though the major proposed methodology sounds not novel enough, the paper also included dive deep analysis of the method's effectiveness. The paper conducted analysis on entity embedding norms, entity-specialized attendtion heads, entity-to-entity attention to explore how topology injection affects the model's internal representations and attention behavior which is useful research.

Critical weaknesses

The proposed framework and suggested fine-tuning technique fit into general mechanisms of combining graph embeddings with LLMs which doesn't suggest enough novelty. Besides, the testing datasets none of which include more than millions of events or million-level of nodes and edges are actually small scale in practice. The effectiveness of the method upon real world large scale of data and on complex large graphs is unclear.

Questions for authors

Regarding this claim from the paper "Experiments across enterprise security logs, blockchain transactions, and cybersecurity events demonstrate that SECBFF consistently outperforms text-only and graph-text baselines on classification tasks, with the largest gains in low-context settings such as zero-shot and one-event detection where topological priors are most valuable.", I would suggest to benchmark the fine-tuned model against the latest cutting edge LLMs, i.e. Opus 5.5 etc., to suggest its effectiveness.

Reviewer_NiPiOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies temporal security detection when graph embeddings are available for only a small fraction of the entities in an event stream.

SecBFF projects frozen graph embeddings into an LLM's input space and replaces the corresponding entity-token embeddings in place, preserving token positions and leaving uncovered entities in text form.

The authors report improvements across three datasets and five detection settings, including zero-shot scoring and event localization.

My assessment is based on the manuscript and its reported experiments, without an independent reproduction.

Critical strengths
  1. The main method is explained clearly, particularly the distinction between length-preserving replacement and prefix-based fusion. The appendices provide substantial implementation detail.

  2. The coverage study is particularly informative. Low-coverage training transfers more reliably to changing inference coverage, so robustness to sparse topology should not be equated with robustness to losing topology after training. (Appendix H, pp. 17–18)

  3. The architecture and the principal comparisons support the central contribution. The paper also appropriately acknowledges that the attention-based mechanistic account is associational without head-level interventions. (Appendix L, p. 26)

  4. In-place and length-preserving replacement offers a straightforward way to use structural information while retaining ordinary text processing for uncovered entities (Section 3, pp. 3–4). The explicit preservation of token positions also makes the comparison with the text-only model particularly interpretable.

  5. Tables 1–3 report consistent gains across the three domains; the PR results in Appendix D show that the improvements survive a metric more informative for these imbalanced datasets. Appendix H directly tests coverage scarcity and train–test coverage mismatch, while Random-GFM controls injection placement, coverage, and initial scale. Together these results make a stronger case for the integration method than a single benchmark improvement would.

Critical weaknesses
  1. The chronology and coverage construction need a clearer, reproducible specification.

Sections 4.1–4.2 and Appendix F distinguish graph pretraining from the later LLM data, but do not explicitly give the final LLM training/evaluation split within Elliptic timesteps 35–49 or the later LANL buckets. It is also unclear how the reported 15% coverage is obtained and matched to later entities, especially for LANL's computer–hour nodes, and how ego-graph sequence construction prevents access to information later than the scored event. I appreciate the explicit statement that evaluation sequences are excluded from self-supervised fine-tuning; a concise split/availability specification would substantiate that assurance and strengthen the temporal-detection claims.

Questions for authors
  1. Clarification: Could you provide the exact downstream train/evaluation partitions for Elliptic and LANL, together with the rule that makes a later entity graph-covered?

  2. Clarification: What is randomized in the three zero-shot runs in Tables 1–3? Is it the self-supervised training, the data construction, or another stage?

Reviewer_QjY3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SecBFF, a framework for incorporating graph-derived entity representations into LLMs for security event analysis under partial graph coverage. Its designs a Topology-Injected Masked Fine-Tuning method which replaces covered entity mentions with projected graph embeddings while retaining textual representations for uncovered entities. Experiments across three datasets and five detection settings report improvements over baselines.

Critical strengths
  1. The motivation to incorporate entity interaction structure into security event analysis under incomplete graph coverage is clear and practically relevant.

  2. The proposed integration is easy to follow. It preserves the original vocabulary and sequence length while accommodating entities with and without graph embeddings.

  3. The evaluation covers three datasets and five detection settings, with consistently better results and additional analyses of embedding coverage and model behavior.

Critical weaknesses
  1. The necessity of incorporating the LLM into such task is not convincingly demonstrated. The proposed projection and entity-replacement mechanism does not appear to depend on capabilities specific to LLMs and could also be applied to smaller language models or other sequence models. Moreover, assigning zero-vector embeddings to uncovered entities (Appendix E) may substantially impair the graph-based baselines, raising concerns about comparison fairness.

  2. The graph embeddings are precomputed general-purpose representations and are not guaranteed to preserve task-relevant information in entity names or descriptions. However, for covered entities, the proposed method replaces their textual representations entirely with these embeddings. The potential loss of complementary textual information and its implications for detection are insufficiently discussed.

  3. The ablations do not clearly isolate the contributions of the main design choices. The prefix baselines differ in both injection strategy and graph aggregation, obscuring the benefit of entity replacement. The effectiveness of the linear alignment and the respective contributions and potential trade-offs of jointly training the projector and LoRA also remain unclear.

  4. The framework relies on precomputed graph embeddings, but their reliability under incomplete or evolving graphs remains unclear. The coverage study varies embedding availability rather than the completeness or freshness of the underlying graph. How does the method perform when embeddings are constructed from missing interactions or become outdated as entity relationships change?

Questions for authors

It would be good to provide a controlled comparison with a non-LLM sequence model that receives the same textual inputs and graph embeddings under the same conditions. This would help clarify whether SecBFF’s gains come from using a LLM specifically or from combining textual and graph information.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.