Counterfactual Security Verification: Toward Robust Policy Attribution and Cross-Domain Generalization

Security verification systems are essential for evaluating the effectiveness of defense policies in networked environments. However, existing approaches often lack causal interpretability and struggle to generalize across heterogeneous deployment scenarios. In this paper, we propose a causalityaware verification framework that enhances both attribution and generalization in policy evaluation. Our method combines deep representation learning with structural causal modeling to capture the interventional effects of policy decisions under diverse attack behaviors and contextual conditions. A counterfactual inference module enables fine-grained diagnosis of policy failures, while invariant risk minimization ensures transferability across unseen environments by identifying stable, causally relevant features. We evaluate our approach on a realistic network security simulation platform using the CICIDS2017 dataset. Experimental results show that our method improves attribution precision by over 13% and cross-domain AUC by over 10% compared to prior baselines, demonstrating its effectiveness in robust and interpretable security policy verification.