[論文レビュー] Adversarial Network Imagination: Causal LLMs and Digital Twins for Proactive Telecom Mitigation

Vignesh Sriram, Meng Yu+2 more|arXiv (Cornell University)|Jan 9, 2026

Software-Defined Networks and 5G被引用数 0

ひとこと要約

要約: The paper presents a closed-loop framework that uses a Causal LLM, a Knowledge Graph, and a Digital Twin to generate, simulate, and evaluate adversarial telecom failure scenarios for proactive mitigation.

ABSTRACT

Telecommunication networks experience complex failures such as fiber cuts, traffic overloads, and cascading outages. Existing monitoring and digital twin systems are largely reactive, detecting failures only after service degradation occurs. We propose Adversarial Network Imagination, a closed-loop framework that integrates a Causal Large Language Model (LLM), a Knowledge Graph, and a Digital Twin to proactively generate, simulate, and evaluate adversarial network failures. The Causal LLM produces structured failure scenarios grounded in network dependencies encoded in the Knowledge Graph. These scenarios are executed within a Digital Twin to measure performance degradation and evaluate mitigation strategies. By iteratively refining scenarios based on simulation feedback, the framework shifts network operations from reactive troubleshooting toward anticipatory resilience analysis.

研究の動機と目的

• Motivate proactive resilience in complex telecom networks by anticipating failures before they occur.
• Integrate causal reasoning, knowledge graphs, and high-fidelity simulation to generate executable failure scenarios.
• Demonstrate that causally constrained generation yields more valid, cascading, and mitigatable scenarios than baselines.
• Evaluate the framework across diverse topologies and adversarial conditions to assess generalizability and mitigation impact.

提案手法

• Define a constrained generation process where a Causal LLM produces failure scenarios grounded in a network Knowledge Graph.
• Ground generation in explicit causal structure to ensure feasible interventions and multi-step reasoning.
• Execute generated scenarios in a Digital Twin to observe propagation and performance impact.
• Use a Mitigation Engine to test recovery actions and feed results back to refine generation in a closed loop.
• Compare with rule-based and replay baselines to assess improvements in scenario diversity and mitigation effectiveness.

実験結果

リサーチクエスチョン

• RQ1 Can a Causal LLM generate causally consistent, executable network failure scenarios under dependency constraints?
• RQ2 Do causal constraints improve the validity, depth of cascading, and mitigation effectiveness of generated scenarios compared to baselines?
• RQ3 Can the closed-loop feedback between simulation and generation uncover more informative adversarial scenarios across different topologies?
• RQ4 How well does the framework generalize to varied telecom topologies (ISP/backbone and AS-level structures)?

主な発見

• Adversarial Network Imagination generates more diverse and complex failure scenarios than baselines, including multi-component and cascading events.
• Causal conditioning enables deeper propagation in the Digital Twin, enhancing stress testing and mitigation gains.
• Ablation shows knowledge graph, causal conditioning, and closed-loop verification each contribute to scenario realism, cascading behavior, and adaptive refinement.
• Cross-topology experiments indicate the Causal LLM adapts to topology-specific dependencies, suggesting broader generalizability.