Explainable Machine Learning Framework Strengthens Intrusion Detection in Next Generation Networks

Researchers evaluating next-generation cellular communication security have demonstrated an explainable ensemble framework that pairs tree-based boosting with deep neural modeling to detect malicious traffic across fifth-generation and beyond networks. In a benchmark evaluation conducted on the public 5G-NIDD dataset, an ensemble combining Extreme Gradient Boosting and a compact multilayer perceptron achieved an overall accuracy of 99.3 percent alongside a 99.2 percent precision rate, a 99.1 percent recall rate, a 99.1 percent F1-score, and an area under the receiver operating characteristic curve of 99.7 percent. Tested against a fixed held-out split of thirty-nine thousand one hundred and thirty network flows, the model maintained a low false-positive rate of 0.9 percent while generating interpretable feature rankings and instance-level decision justifications. The corresponding 95 percent Wilson confidence interval for accuracy spanned from 99.21 percent to 99.38 percent, reflecting finite sampling uncertainty on the specific held-out test distribution rather than variability across random training seeds.

The transition toward fifth-generation architectures introduces fundamental changes to the structure of modern communications by integrating software-defined networking, network function virtualization, cloud-native services, multi-access edge computing, and network slicing. These systems allow multiple isolated logical networks to operate dynamically across shared physical infrastructures, supporting diverse technical applications that range from autonomous transportation and industrial automation to remote healthcare and large-scale Internet of Things ecosystems. However, because next-generation network architectures eliminate fixed perimeter defenses and distribute critical operations across the radio access network, core network functions, edge nodes, and interconnected user devices, the potential attack surface expands significantly. Threats can manifest through slice isolation breaches, virtual function misconfigurations, unauthorized signaling access, or distributed denial-of-service campaigns designed to overwhelm edge resources and disrupt latency-sensitive services.

Defending distributed telecommunications infrastructure relies heavily on automated network intrusion detection systems capable of identifying both known attack signatures and evolving malicious traffic patterns. Although traditional signature-based detection mechanisms reliably catch established threats, they exhibit severe operational blind spots when confronted with polymorphic, zero-day, or slow-rate intrusions. Machine learning and deep neural networks offer flexible alternatives by learning complex statistical relationships directly from network telemetry. Despite their classification power, complex neural architectures function predominantly as uninterpretable black boxes, providing little operational insight into why an alert was triggered. In operational environments, an intrusion alert acts as a critical decision-support signal that can prompt network isolation, connection blocking, forensic scrutiny, or policy adjustments. If an automated system cannot verify its reasoning, administrators risk interrupting mission-critical services over false alarms or allowing silent, low-rate compromises to propagate undetected through edge nodes.

To reconcile high classification accuracy with operational interpretability, the newly evaluated framework implements a multi-stage workflow incorporating data preparation, training-constrained feature engineering, ensemble classification, post-hoc explainability, and graded decision support. Evaluated against the realistic 5G-NIDD benchmark extracted from two base stations connected to the functional 5G Test Network at the University of Oulu, Finland, the underlying data represented normal background communications alongside eight specific attack types: Internet Control Message Protocol flooding, User Datagram Protocol flooding, SYN flooding, Hypertext Transfer Protocol flooding, slow-rate denial-of-service, SYN scanning, Transmission Control Protocol connect scanning, and UDP scanning. The full benchmark distribution contained over four hundred and seventy-seven thousand benign flows, more than four hundred and fifty-seven thousand UDP flood flows, and smaller cohorts of scanning and low-rate attacks, presenting substantial class imbalance.

To guarantee strict experimental fairness and prevent data leakage, the evaluation used a stratified partitioning strategy allocating 70 percent of the data to model training, 15 percent to validation, and 15 percent to final testing under a fixed random seed. Preprocessing routines were fitted strictly on the training partition before being frozen and applied to validation and testing sets. Constant columns, duplicates, explicit network identifiers, sequence-revealing timestamps, and leakage-prone variables were eliminated, while missing numerical values were imputed using training medians. Imbalanced class distributions were mitigated by computing inverse-frequency class weights rather than generating synthetic records. A four-stage feature selection process calculated mutual information scores on the training data, applied recursive feature elimination using an auxiliary tree estimator on the validation set, and selected the smallest feature subset achieving a macro-F1 score within 0.1 percentage point of the optimum.

The core detection layer leverages two distinct model families trained on the exact same selected feature representation to exploit complementary inductive biases. The Extreme Gradient Boosting branch effectively isolates structured, threshold-based relationships and decision boundaries within tabular flow variables, whereas the multilayer perceptron captures dense nonlinear interactions across the normalized feature inputs. The neural branch was configured with three fully connected dense layers of 128, 64, and 32 units, rectified linear unit activations, dropout regularizers of 0.20, and a softmax output, optimized via the Adam algorithm under weighted categorical cross-entropy. Calibrated probability vectors from both branches were combined via weighted probability fusion governed by a scalar parameter alpha. Rather than relying on arbitrary manual tuning, alpha was optimized across validation grid points to maximize validation macro-F1 while curbing false alarms, establishing an optimal operating value of 0.6 that favored the tree classifier while preserving dense neural representations.

Interpretable decision support is integrated through dual explainable artificial intelligence techniques that assign quantifiable importance to traffic indicators. Global feature importance was computed using SHapley Additive exPlanations grounded in cooperative game theory, establishing that the ensemble model’s global classification behavior relied primarily on packet transmission rates, flow durations, SYN flag behavior, destination-port entropy, source byte volumes, and connection states. Complementing these macroscopic patterns, Local Interpretable Model-Agnostic Explanations were deployed to construct local surrogate approximations around individual flow instances. This localized validation allows security analysts to inspect precisely which packet counts or port anomalies contributed to an individual alert, distinguishing genuine malicious behaviors from transient network anomalies or dataset artifacts.

Statistical evaluation on the held-out sample confirmed that the ensemble model outperformed a diverse collection of standalone machine learning and deep learning baselines, including single Decision Trees, Random Forests, standalone XGBoost, standalone Multilayer Perceptrons, Long Short-Term Memory networks, and hybrid Convolutional Neural Network and Long Short-Term Memory models. Paired two-proportion z-tests with Holm correction established that the accuracy gain of the proposed ensemble over each baseline remained statistically significant at an adjusted p-value below 0.001. Compared with the strongest baseline model—the convolutional-recurrent network which reached 98.7 percent accuracy—the proposed framework achieved a 0.6 percentage-point accuracy increase, representing an approximate 46.2 percent relative reduction in total classification errors across the thirty-nine thousand one hundred and thirty test flows.

Ablation experiments demonstrated that each framework component directly supported the observed performance gains. Excluding feature selection produced the sharpest increase in false-positive rates, while removing inverse-frequency class weighting substantially reduced recall on minority attack categories. Standalone XGBoost and standalone multilayer perceptron branches underperformed the fused configuration, and unweighted probability averaging also lagged behind validation-optimized weighting. While post-hoc SHAP and LIME components do not modify classification outputs, quantitative evaluation verified surrogate fidelity and stability under label-preserving input perturbations constrained to valid training ranges, ensuring that local explanations remained consistent across similar traffic samples.

Computational latency profiling revealed distinct timing characteristics across the operational pipeline, providing a clear rationale for a bifurcated, two-path deployment architecture. On the evaluated workstation platform, raw detection required an average of 22.6 milliseconds per flow after warm-up, yielding a serial-equivalent throughput of approximately 44.2 flows per second. In contrast, calculating SHAP feature attributions required an additional 26.5 milliseconds, while generating local LIME explanations required an additional 34.2 milliseconds, expanding total end-to-end processing time to 83.3 milliseconds per flow for a fully explained alert. Because executing dual explainability routines for every packet stream would overburden edge computing nodes, the investigators outlined a selective execution strategy where benign, high-confidence flows follow the rapid 22.6-millisecond edge path, while full explanations are generated selectively for high-risk, ambiguous, or policy-sensitive incidents requiring operator confirmation.

Inspection of the residual error patterns documented in the held-out confusion matrix revealed that misclassifications were concentrated among attacks exhibiting overlapping temporal and volumetric characteristics. High-rate volumetric floods were detected with exceptional reliability, whereas slow-rate denial-of-service attacks showed slightly lower F1-scores due to behavioral overlap with normal connection activity and HTTP-based attacks. To address these operational ambiguities, the framework links predicted attack probabilities directly to four graded response actions: allowing normal traffic, monitoring uncertain connections, restricting suspect sources, and automatically blocking confirmed high-risk threats, ensuring that low-agreement alerts trigger monitoring rather than disruptive automated shutdowns.

The study clearly identifies several foundational boundaries and methodological constraints. Because the experimental results originate from a single public benchmark, the findings cannot guarantee cross-operator generalizability across diverse commercial topologies, proprietary encryptions, or dynamic traffic policies without external field validation. Furthermore, the documented performance metrics reflect a fixed held-out data partition without quantifying random-seed variance or multi-site dispersion, and reliance on supervised labels inherently constrains sensitivity against novel zero-day exploits. The authors conclude that practical deployment within resource-constrained multi-access edge computing nodes requires rigorous platform-specific resource profiling, model compression, horizontal detector replication, and longitudinal validation across diverse telecommunication environments.

Subject of Research: Explainable ensemble machine learning for 5G network intrusion detection

Article Title: An explainable XGBoost–MLP fusion framework for intrusion detection in 5G and beyond networks

Article References: Al-Balasmeh, H. (2026). An explainable XGBoost–MLP fusion framework for intrusion detection in 5G and beyond networks. Discover Artificial Intelligence, 6(1), Article 1573. https://doi.org/10.1007/s44163-026-02371-0

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02371-0

Keywords: network intrusion detection, 5G security, explainable AI, XGBoost, multilayer perceptron, SHAP, LIME, machine learning, explainable, fusion, framework, intrusion