Trustworthy, Explainable, and Sustainable Decentralized Intelligence for 6G Networks

Summary

This is a position/vision article (submitted to IEEE Signal Processing Magazine) arguing that AI-native 6G security requires jointly designing decentralization, trustworthiness, explainability, and sustainability rather than treating them separately. The authors argue centralized ML pipelines fail 6G on scalability, latency, and multi-stakeholder governance grounds, and formalize decentralized federated learning (DFL) through a graph-regularized consensus objective (Eq. 1-2) whose robust-aggregation interface can incorporate peer trust, uncertainty, explanations, and energy state. They survey lifecycle threats (poisoning, adversarial inputs) and defenses (robust aggregation, adversarial training, HE/SMC/TEE/differential privacy), then propose that feature attributions plus calibrated uncertainty should serve as machine-readable control-plane signals driving IDS self-optimization, concept-drift detection, automated O-RAN (xApp/rApp) responses, and cross-peer alert verification — while warning that explainers themselves become attack targets (fairwashing, backdoored explanations). Two explicit untested hypotheses are stated: attribution-distribution shift as an early concept-drift detector, and the same signal to expose explainer tampering. On sustainability the paper advocates neuromorphic hardware (SNNs, reservoir computing, citing up to three-orders-of-magnitude EDP gains on Intel Loihi) and joint energy-/age-of-information-aware participation scheduling for energy-harvesting nodes. Limitations: the work is entirely conceptual — no experiments, algorithms, or quantitative evaluation; only 12 references support claims spanning four large fields; the unified design remains an assertion rather than a mechanism.