acceptodds
Under review as a conference paper at ICLR 2027

GraphGuard: Graph-Constrained Neuro-Symbolic Synthesis of Network Defense Configurations

Abstract

Translating high-level intents into executable configurations is essential for modern network defense. Existing Large Language Model (LLM)- and Retrieval-Augmented Generation-based methods still rely heavily on unconstrained LLM reasoning to assemble configuration fields and instantiate runtime parameters, making reliable configuration synthesis difficult. This difficulty stems from fragmented configuration knowledge across heterogeneous devices, interdependent configuration choices tied to dynamic network contexts, and generated configurations that may violate deterministic constraints or deployment requirements. To address these issues, we propose GraphGuard, a graph-constrained neuro-symbolic framework that translates high-level defense intents into verified configurations for heterogeneous devices. GraphGuard organizes configuration knowledge into a unified typed five-slot space, uses graph-constrained joint synthesis to compose compatible configuration components, and deterministically instantiates runtime-dependent parameters under explicit operational constraints. A structured verifier then guides bounded repair through graph-defined legal edits, while a backend-independent Intermediate Representation (IR) enables device-specific compilation. Experiments in heterogeneous DDoS mitigation environments show that GraphGuard achieves 93.45% Joint Exact Match and 75.41% Safe Synthesis Success, with a 1.50% unsafe acceptance rate on infeasible instances, while maintaining practical end-to-end synthesis latency.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Network Configuration Synthesis, Neuro-Symbolic Reasoning, Knowledge Graphs, Intent-Based Networking, Constraint Solving

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

The paper proposes GraphGuard, a neuro-symbolic framework for translating natural-language network-defense intents into device configurations. It combines a typed five-slot configuration graph, learned joint structural scoring, deterministic parameter solving, bounded repair, and backend compilation through an intermediate representation. Experiments cover overall synthesis quality, component ablations, compositional and linguistic generalization, runtime perturbations, search budgets, and backend portability. GraphGuard improves safe synthesis success over the evaluated retrieval, agentic, and symbolic baselines.

Critical strengths
  • The paper addresses a well-motivated problem in which natural-language interpretation must be combined with strict operational constraints. Its formulation captures dependencies among configuration choices, runtime resources, policies, and backend capabilities, providing a concrete setting for studying the interaction between learned semantic interpretation and symbolic reasoning.

  • The framework assigns clear responsibilities to its components. Learned scoring supports intent alignment and joint structural selection, while deterministic solving handles numerical feasibility. The typed configuration representation connects these stages and provides a consistent basis for checking dependencies, conflicts, resource limits, and backend realizability.

  • Verification, repair, and compilation are integrated into a coherent release procedure. Repairs are restricted to enumerated legal edits and followed by rechecking, while the intermediate representation supports explicit checks of values, units, scope, and lifecycle across backends. The paper also clearly states the conditions under which these checks provide guarantees.

  • The empirical evaluation is comparatively broad and goes beyond a single overall accuracy table. It includes agentic and symbolic baselines, extensive component ablations, distribution shifts, runtime counterfactuals, and budget analysis. Separating unsafe acceptance, over-refusal, structural correctness, and conditional executability helps reveal trade-offs that would be obscured by reporting success rate alone.

Critical weaknesses
  • The Symbolic-Only baseline provides limited evidence about the advantage over strong symbolic synthesis. Its compositional-OOD performance falls to zero because it uses exact-motif lookup, whereas a symbolic system could potentially recombine known prototypes under the same constraints. Since compositional generalization is a central claimed benefit, a comparison using the same candidates, solver, and verifier with deterministic structural ranking would more directly establish the contribution of learned joint scoring.

  • The additional value of LLM-guided repair is not convincingly established. On the initially invalid proposal pool, its recovery rate is 7.7%, compared with 7.1% for severity-greedy selection under the same attempt limit. This small conditional difference needs uncertainty and cost analysis, particularly because LLM selection introduces additional inference calls. The ablation supports the usefulness of bounded repair as a whole, but does not establish that its decision policy needs an LLM.

  • Some safety-related ablation results require a clearer causal explanation. Removing hard search constraints increases unsafe acceptance from 1.50% to 12.30%, although the intended final verifier retains the relevant checks. A downstream verifier should still reject violations of constraints it checks correctly. A stage-level breakdown is needed to explain whether these additional unsafe releases arise from different constraint coverage, semantic selection errors, or changes in the release path, and to distinguish the effects of pruning from verification.

  • The practical safety evidence depends heavily on the correctness and coverage of the frozen configuration graph and backend contracts. The experiments examine runtime and constraint changes, but do not establish robustness to missing dependencies, incorrect compatibility relations, or inaccurate resource bounds. Backend acceptance and round-trip consistency provide useful evidence of configuration preservation, while the reliability of the underlying knowledge remains insufficiently evaluated. This limits how broadly the reported safety results can be interpreted.

Questions for authors
  1. How would a compositional symbolic baseline perform when given the same candidate prototypes, hard constraints, numerical solver, and final verifier, with deterministic ranking replacing learned structural scores? This would help assess the contribution of neural scoring beyond exact-motif lookup.

  2. What explains the increase in unsafe acceptance when search-stage hard constraints are removed while the final verifier is retained? Please identify the violated conditions and the stage at which the additional unsafe configurations pass through the system.

  3. Does LLM-guided repair improve final safe synthesis success over severity-greedy repair under comparable computational budgets? Please report the corresponding inference cost, latency, and uncertainty on the same invalid proposal pool.

  4. What evidence establishes the accuracy and coverage of the constructed configuration graph? How sensitive are the results to a missing dependency, an incorrect relation, or an inaccurate numerical bound? Please also identify the actual graph and model configurations used to produce the reported results.

3 comments
AuthorsOpenReview ↗

Response to Reviewer JwqH

We thank Reviewer JwqH for the insightful review and for recognizing our problem formulation and the breadth of our empirical evaluation. We have incorporated your suggested analyses into the revised manuscript:

  1. Compositional Symbolic Baseline (Comp. Symbolic, Question 1):

    • Following your specification, we implemented Comp. Symbolic in Table 1 (RQ1), Figure 3a (RQ3), and Appendix E.1 (Table 6). It uses the exact same candidate sets Kq,t\mathcal{K}_{q,t}, hard predicates Ωq\Omega_q (B=8B=8), CP-SAT solver, and verifier Jver\mathcal{J}_{\mathrm{ver}} as GraphGuard, but replaces learned scoring EΘE_\Theta with deterministic BM25 lexical similarity.
    • Insight: On Compositional-OOD (Figure 3a), while exact-motif lookup (Symbolic-Only) drops to 0.0%, Comp. Symbolic achieves 76.20% JEM and 58.86% SSS, proving our five-slot graph representation and hard constraints Ωq\Omega_q enable structural recombination. Crucially, GraphGuard's learned neural potentials EΘE_\Theta provide an additional +15.31% JEM and +14.20% SSS gain on Compositional-OOD (and +10.05% JEM / +8.56% SSS on IID), showing the value of learned relational scoring.
  2. Stage-Level Breakdown of Search Pruning vs. Verification (Question 2):

    • We separated w/o Search Pruning from w/o Hard Constraints in Table 2 (RQ2) and Appendix E.1:
    • w/o Search Pruning (disabling Ωq\Omega_q during beam search while keeping the full downstream verifier Jver\mathcal{J}_{\mathrm{ver}}): Search pruning drops to 0.00% and verifier rejection jumps to 58.40%. Because the verifier still enforces relational rules, UAR remains statistically unchanged (1.65% vs. 1.50%), confirming that the verifier blocks relational violations. However, unpruned conflicting prefixes saturate the beam (B=8B=8) and crowd out valid motifs, causing ORR to nearly triple (from 4.80% to 14.20%), JEM to drop by 9.25 points, and local P95 latency to double (4.610 ms).
  3. Cost, Latency, and Semantic Value of LLM-Guided vs. Severity-Greedy Repair (Question 3):

    • We evaluated repair policies on the invalid pool (Ninv=372N_{\mathrm{inv}}=372, 10.33% of outcomes) in Section 5.6 (RQ5, Figure 5b) and Appendix E.4:
    • While deterministic Severity-Greedy restores constraint feasibility on 68.70% of cases, it greedily minimizes violation weight without intent awareness, achieving 51.20% conditional semantic SSS. In contrast, LLM-Guided repair achieves 68.40% conditional semantic SSS (+17.20% relative improvement, p<0.001p < 0.001) while requiring 85.81% fewer solver/verifier calls per invalid case (2.10 vs. 14.80 calls).
  4. Configuration Graph Accuracy and Corruption Sensitivity (Question 4):

    • Extraction Accuracy (Appendix A.3): Built from sanitized manuals and rules, Gcfg\mathcal{G}_{\mathrm{cfg}} contains 442 canonical prototypes and 618 typed edges. Our extractor McfgM_{\mathrm{cfg}} achieves 93.7% Prototype F1 and 91.8% Relation F1 on held-out segments.
    • Corruption Sensitivity (Section 5.5, Figure 4b & Appendix E.3): Injecting 5%–15% synthetic corruptions into Gcfg\mathcal{G}_{\mathrm{cfg}} shows that missing dependencies primarily increase conservative abstention (14.60% ORR) rather than unsafe execution. Even under 15% false compatibility or bound shifts, our layered runtime and backend gates cap worst-case UAR at 7.50%.

We have provided the policy-permitted core algorithm-related code and experimental records in the supplementary package for your review. If our clarification and revision address your concerns, we hope you will carefully review the content of our paper again. We would be most grateful if you could consider raising the score.

Thank you so much for writing the response, I will retain my score as 3.

AuthorsOpenReview ↗

Thank You for Your Support and Constructive Review

Dear Reviewer JwqH,

Thank you very much for taking the time to read our response and revised manuscript, and for your positive evaluation and encouraging support of our work. Your constructive feedback has significantly helped us strengthen the clarity and rigor of the paper.

Best regards,
The Authors

Reviewer_ZZ7kOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces GraphGuard, which translates network defense requirements into device configurations. The author combines several components as workflow to process data, and use predefined rules to ensure the reliability of the workflow.

Critical strengths
  1. This paper is for a specific domain, which try to reduce the unstable when machine generates data.
  2. GraphGuard organizes heterogeneous configuration knowledge into Match, Action, Scope, Cost, and Lifecycle, providing a common configuration space across different devices. The proposed method has more value in engineering applications in this field.
Critical weaknesses
  1. “Example material, pending replacement with the final implementation and experimental records” in A5, D, and F.3, etc. The attachments submitted are raw drafts. Not sure if your paper is AI-generated, but this is an incomplete manuscript, and I doubt whether other parts of your paper have been checked.
  2. The workflow idea is well explored by many papers. The combination of components is not counted as a good contribution. For the specific domain, the workflow may be more advanced than current practice, but it is not sufficient.
  3. The predefined rules, etc., are heuristic, which is weak for guaranteeing reliability.
Questions for authors
  1. Please clearly explain what new capabilities are gained through the integration.
  2. 4.1 Describe the flow to prepare the dataset. Did you try to extract documents from data sources? Were your experiments conducted on your extracted datasets?
  3. The appendix says that the examples are not real, so what is your real information for the paper?
2 comments
AuthorsOpenReview ↗

Response to Reviewer ZZ7k

We thank Reviewer ZZ7k for reviewing our manuscript and recognizing the engineering value of our five-slot representation. We address your concerns regarding the appendix presentation and technical contributions below:

  1. Appendix Presentation, NDA Constraints & Manuscript Completeness:

    • Appendix Presentation & Enterprise NDA: We apologize for the editing oversight that left internal sanitization tags ("Example material, pending replacement...") in parts of the initial Appendix (A.5, D, and F.3). Due to enterprise security compliance regulations and vendor non-disclosure agreements (NDAs) governing proprietary device manuals (Huawei SecoManager and NSFOCUS ADS), raw internal files cannot be publicly redistributed; thus we used sanitized canonical illustrations in the Appendix. In the revised manuscript, we cleaned up residual tags, added a Data Confidentiality & Sanitization Statement at the beginning of Appendix A, and provided hyperparameters in Appendix D.5 (Table 5).
    • Main Manuscript Completeness: The illustrative formatting slip was confined to the supplementary Appendix. The 9-page main text (Sections 1–6) is self-contained, rigorously presenting our formal definitions, graph-coupled search, CP-SAT optimization, and RQ1–RQ7 evaluations. Under ICLR guidelines, appendices provide supplementary context, and we request evaluation based on the main text.
  2. Dataset Preparation and Knowledge Graph Evaluation:

    • As detailed in Section 4.1 and Appendix A.3, experiments were conducted on the extracted configuration graph Gcfg\mathcal{G}_{\mathrm{cfg}} (∣V∣=442|\mathcal{V}|=442 canonical prototypes, ∣E∣=618|\mathcal{E}|=618 typed edges). Evaluated on 200 held-out segments, our extractor McfgM_{\mathrm{cfg}} achieves 93.7% Prototype F1 and 91.8% Relation F1.
  3. Capabilities & Guarantees:

    • Capabilities (Section 5, Tables 1–2): GraphGuard combines neural intent scoring with symbolic constraint solving, achieving (i) 73.06% Compositional-OOD SSS via motif recombination, (ii) 64.20% in-search pruning to eliminate structural conflicts early (reducing P95 latency by 41.4%), and (iii) 0.0053 NPE via CP-SAT numerical optimization.
    • Formal Guarantees (Appendix B.2–B.4): We formalize the release gate BackendOK⁡\operatorname{BackendOK} and prove two properties: Property 1 (Checked Release) ensures released payloads satisfy verifier predicates and backend contracts; Property 2 (Bounded Repair Termination) ensures repair decreases violation severity and terminates in at most Tmax⁡=3T_{\max}=3 rounds.

We have provided the policy-permitted core algorithm-related code and experimental records in the supplementary package for your review. If our clarification and revision address your concerns, we hope you will carefully review the content of our paper again. We would be most grateful if you could consider raising the score.

Thank you for your detailed responses and clarifications. While the revisions help clarify the appendix issues and the proposed technical approach, I still have concerns about the methodological novelty, the reliability of the predefined rules, and the completeness of the experimental documentation. Therefore, I will maintain my original rating.

Reviewer_yqUFOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The paper addresses translating natural-language defense intents into executable, verified configurations for heterogeneous devices. Its main contribution is a graph-constrained neuro-symbolic framework that combines a unified typed configuration representation with runtime-aware parameter solving, bounded repair, and verified backend compilation. The proposed GraphGuard organizes configuration knowledge into a typed five-slot knowledge graph (Match/Action/Scope/Cost/Lifecycle). At inference time, it retrieves candidates, selects structurally consistent configurations through typed beam search, solves numerical parameters using CP-SAT, and applies deterministic verification and bounded graph-defined repair. Verified configurations are compiled into device-specific payloads through a backend-independent IR; the system abstains if no candidate passes the required checks within its search and repair budgets. Experiments on a self-built DDoS-mitigation benchmark show improved configuration accuracy and safety, with additional analyses of component contributions, generalization, robustness, and backend portability.

Critical strengths
  1. The paper addresses an important practical problem. Translating defense intents into valid configurations is challenging because device-specific knowledge is fragmented, configuration choices depend on runtime conditions, and errors can disrupt critical services. The framework provides a structured approach to handling these dependencies across heterogeneous devices.

  2. The reported results demonstrate strong overall performance. Component ablations support the value of typed graph reasoning and numerical solving, and experiments across three language-model backbones suggest that the gains are not restricted to one backbone.

Critical weaknesses
  1. Incomplete experimental records limit reproducibility and verification of the main results. Appendices A.5, D, and F.3 contain material pending replacement with final implementation and experimental records; backend timing units remain unresolved, and the main comparisons lack uncertainty estimates. Evaluation is also restricted to a self-built benchmark, without validation on an external benchmark such as NetConfEval, where applicable, or independently sourced defense-configuration cases, limiting reproducibility and confidence in generalization beyond the constructed setting.

  2. The correctness of the configuration knowledge graph is insufficiently validated. The framework relies on the frozen graph G_cfg, but the paper reports neither extraction-accuracy measurements nor sensitivity tests for incorrect or missing relations. Schema checks establish well-formedness rather than factual correctness, leaving the effect of knowledge errors on end-to-end safety unclear.

Questions for authors
  1. The paper builds on generation-and-constraint-solving paradigms such as Synchromesh and ConstraintLLM. Could the authors include a baseline from this family, beyond retrieval-based and ReAct-style methods, to clarify whether the gains come from coupling the typed configuration space, runtime parameter solving, and verified compilation, or simply from adding a solver and verifier?

  2. The framework relies on the correctness of the configuration knowledge graph, G_cfg. Could the authors provide extraction-quality measurements, such as precision and recall, or experiments testing sensitivity to incorrect or missing graph knowledge? This would strengthen confidence in the safety provided by the verifier.

  3. The AI Use Statement says that details are described in the paper, but these are not clearly identified. Please provide the relevant details or specify where they can be found.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.