Abstract
With the rapid development of 6G networks, anomaly detection in network edge intelligence faces significant challenges in system interpretability and trustworthiness. Although machine learning-based methods improve detection performance, their black-box nature limits reliable cybersecurity decision support. To address this, we propose a novel framework integrating causal inference with LSTM networks. Our approach first applies Random Fourier Feature transformation to eliminate nonlinear feature correlations-a prerequisite for valid causal analysis. We then quantify feature-specific causal effects using sample-weighted adjustments to ensure model stability. Furthermore, Generative Adversarial Networks generate high-quality minority-class samples to augment training data, enhancing anomaly detection accuracy. Experimental validation on two large-scale datasets demonstrates a 33.7% improvement in explainability and a 68% reduction in root-cause localization time. This work establishes a new cybersecurity paradigm for 6G edge intelligence through causal reasoning.