Explainable AI Could Make or Break Trust in 5G and 6G Networks

Artificial intelligence now sits at the heart of 5G networks, deciding how spectrum is shared, how beams are steered, and how intrusions are flagged. But the models behind these decisions are often opaque, and a new scoping review argues that the explanations meant to illuminate them may themselves carry hidden risks. Published in Discover Internet of Things, the review maps a fragmented research landscape in which explanations can be unstable, unfaithful, or even leak private data, and it proposes a trade-off-aware framework for deciding when an explanation is actually worth its cost.

After title-and-abstract screening removed 324 records and full-text assessment excluded 197 more, 73 studies made the final corpus. The authors chose a scoping design rather than a meta-analysis because the literature spans heterogeneous wireless tasks, datasets, models, explanation methods, and deployment environments that resist statistical pooling.

At the core of the paper is a conceptual distinction between transparency and interpretability. Transparency concerns whether a system reveals its inputs, structure, and processing logic; interpretability asks whether a human stakeholder can actually understand and use that disclosure. A raw parameter dump may be transparent in a narrow technical sense yet operationally useless. The authors extend this with two new framing tools, marginal transparency and marginal interpretability, which treat explanation depth as a variable with diminishing returns. As disclosure complexity grows, so do latency overhead, privacy exposure, and cognitive burden, and the review presents an illustrative net-utility equation balancing these costs. The authors stress this is a design heuristic, not a validated optimization law.

The methodological landscape is dominated by two model-agnostic techniques. LIME builds simple surrogate models around individual predictions by perturbing inputs, such as reported signal-to-noise ratio or mobility patterns, and fitting a small linear model to the local decision boundary. This can show an operator that a handover was triggered mainly by declining SINR rather than sudden cell congestion. But the review warns that LIME’s stability depends heavily on how perturbations are designed, and that random perturbation procedures can produce unstable explanations across runs, a serious problem in highly dynamic, non-linear wireless environments.

SHAP, by contrast, draws on game theory, treating each feature as a player contributing to the model’s decision. In 5G core intrusion detection, SHAP can reveal that abnormal port activity consistently carries high importance while mobility features matter less. Its strength is consistency; its weakness is computational cost. Exact Shapley value computation requires evaluating the model across all possible feature subsets, growing exponentially with feature dimension. In large-scale 6G systems with high-dimensional channel states and traffic statistics, this becomes infeasible for real-time operation. Approximations such as KernelSHAP reduce overhead but introduce noisier, less faithful attributions, turning the choice of approximation into a system-level design decision rather than a purely algorithmic one.

The review also examines inherently interpretable glass-box models. Decision trees remain popular where clarity is critical, though shallow trees sacrifice accuracy and deep trees sacrifice readability. Attention mechanisms embedded in neural networks can highlight which antennas were decisive in beamforming decisions. Large language models occupy a more dangerous dual role: as explanation generators translating attributions and telemetry into operator-facing narratives, and as decision-support agents that may recommend live network changes. The authors caution that fluent natural-language output is not evidence of faithfulness, and that operational agents face hallucinated commands, prompt injection embedded in logs, and poisoning of telemetry streams. They propose a safer workflow of verified retrieval, grounded generation, evidence-linked explanation, policy verification, sandbox validation, human authorization, and controlled execution.

Privacy emerges as a central tension. Wireless systems carry location traces, mobility patterns, and device identifiers, and explanations themselves can leak this data, for example by revealing that a specific user’s movement pattern triggered a handover. Differentially private explanations add controlled noise to protect identities, while federated explainability generates explanations locally at edge devices without sharing raw data. But stronger privacy protection often degrades explanation fidelity, and overly aggressive privacy budgets may render explanations too coarse for effective debugging or decision-making in time-critical operations.

One application example illustrates the promise. Each device performs local training and generates attention-based explanations internally, so raw data and sensitive attribution signals never leave the device. The review is careful to note that attention weights should not automatically be treated as causal or faithful explanations without additional validation, and that direct transfer to operational 5G and 6G networks requires further validation.

Adversarial robustness adds another layer of complexity. When these negative regions expand or relocate, it signals internal model degradation even before misclassification rates rise. Yet explanations themselves can be manipulated by sophisticated attackers, and a system can retain accuracy while producing a manipulated attribution map. The review argues that prediction robustness and explanation robustness must be tested separately.

The cross-study synthesis identifies four recurring patterns: evaluation remains fragmented, with many papers reporting predictive performance and visual explanations without jointly measuring fidelity, stability, human usefulness, and latency; deployment evidence is limited because simulations, legacy datasets, and controlled testbeds dominate; explanation criteria are modality-dependent, so I/Q saliency, flow-feature attribution, and natural-language summaries cannot be assessed by a single universal metric; and privacy leakage, adversarial manipulation, uncertainty, and calibrated human reliance are rarely evaluated together. The authors propose a research roadmap running from standardized benchmarks and low-latency explanations in the short term, through integration with Zero Trust architectures and federated edge intelligence in the mid-term, toward self-explaining 6G systems and global standardization by 2030. They emphasize that XAI should not be assumed to create trust automatically; its value depends on whether an explanation is faithful, timely, privacy-aware, actionable, and appropriate for the target network plane and stakeholder.

Subject of Research: Explainable artificial intelligence for security, privacy, and governance in 5G and emerging 6G wireless networks

Article Title: Explainable artificial intelligence for secure and privacy-preserving 5 G networks

Article References: Liao, Q., Chen, Y., He, S., Wang, R., Xu, W., Shen, H., & Chu, W. (2026). Explainable artificial intelligence for secure and privacy-preserving 5 G networks. Discover Internet of Things, 6(1), Article 190. https://doi.org/10.1007/s43926-026-00499-0

Image Credits: AI Generated

DOI: 10.1007/s43926-026-00499-0

Keywords: explainable AI, 5G, 6G, wireless networks, LIME, SHAP, privacy-preserving explainability, Open RAN, network slicing, Zero Trust, federated learning, large language models