X-IDS: explainable intrusion detection for mobile edge computing
Аннотация
Mobile Edge Computing (MEC) has emerged as a critical infrastructure for latency-sensitive applications; however, its distributed and resource-constrained architecture significantly increases exposure to sophisticated cyber threats. Among these threats, Distributed Denial of Service (DDoS) attacks remain particularly destructive due to their ability to exhaust limited edge resources in real time. To counter such threats, recent advances in Deep Learning (DL) have demonstrated strong detection performance. Despite their high accuracy, these models operate as opaque black boxes, limiting trust, auditability, and regulatory compliance in mission-critical MEC environments. This paper proposes X-IDS, an explainable intrusion detection framework that bridges the gap between high detection performance and interpretability. Specifically, a robust AE–MLP baseline (Expert Model) is reproduced using the large-scale NF-UQ-NIDS-v2 dataset. To address the interpretability limitation, a latent-space-driven synthetic distillation strategy is introduced. MiniBatch KMeans clustering is applied within the 16-dimensional latent space to capture intrinsic traffic distributions. Based on cluster statistics, 80,000 synthetic samples are generated and labeled by the trained AE–MLP Expert Model to construct a distillation dataset. The Analyst Model (Logistic Regression) is then trained on this synthetic data. The Expert Model achieved 99.15% accuracy on an independent 100,000-sample test set. The X-IDS Analyst Model, evaluated on the same independent test set, achieved 95.92% accuracy with 96.51% fidelity to the Expert Model while reducing end-to-end inference latency by 49.80% (from 301.78 to 151.49 ms for 5000 flows, averaged over 20 runs). The proposed X-IDS framework enables high-fidelity intrusion detection with global interpretability, making it suitable for real-time deployment in resource-constrained MEC environments.