XAI-Driven and Privacy-Preserving Federated Learning for Secure Intrusion Detection in 6G Fog-IoT Environments

Abstract

The convergence of 6G radio access, software-defined fog infrastructure, and large-scale IoT deployments is reshaping a long-standing problem in network and service management: detecting intrusions close to where traffic originates, without exporting raw telemetry from the fog tier, and without turning the detector into an opaque decision-maker. This paper proposes XAI-FedFog-HybridNet, a federated intrusion-detection architecture for 6G fog-IoT environments that combines a three-branch Bi-LSTM-RNN-CNN classifier, differential-privacy-bounded secure aggregation with Byzantine-robust selection, a permissioned blockchain audit layer, and a dual explainability module built on SHAP and Grad-CAM. Rather than treating these four mechanisms as independent add-ons, the design couples them directly: the privacy budget interacts with the robust-aggregation rule, the blockchain ledger stores explanation digests rather than raw attribution vectors so that auditability does not itself create a privacy leak, and explanations are computed on the aggregated global model rather than on individual client updates. This paper presents the architecture, the federated training protocol, the audit and explainability design, the implementation plan (testbed topology, software stack, and hyperparameter schedule), and an evaluation on two real public network-traffic datasets, IoT-23 and ToN-IoT, using a single-node proxy of the detector (a multi-layer perceptron standing in for the full hybrid architecture). Results are mixed and reported without adjustment: the proxy is strong on ToN-IoT (99.2% binary, 96.8% multiclass accuracy) but is consistently outperformed by a Random Forest baseline on IoT-23, a finding discussed in Section 5-6 as motivation for implementing the full architecture rather than relying on the simplified proxy. The federated, differential-privacy, and blockchain layers are described architecturally but not exercised in this single-node evaluation.