The rapid integration of Industrial Internet of Things (IIoT), cyber-physical systems, intelligent sensors, and connected manufacturing platforms has created highly distributed industrial environments in which cybersecurity must operate under stringent requirements for privacy, low latency, interpretability, and operational continuity. Conventional centralized machine-learning-based intrusion detection requires the aggregation of potentially sensitive industrial data, creating additional privacy and data-governance risks. Federated learning (FL) provides an alternative paradigm in which participating industrial nodes collaboratively train a shared model without directly exchanging their raw observations. However, privacy preservation alone is insufficient for industrial cybersecurity because security operators must understand why a model has classified an industrial event as malicious or anomalous. This paper develops a research-oriented framework for explainable federated learning (XFL) for privacy-preserving cybersecurity in IIoT and smart-factory environments. The proposed methodology integrates distributed model training, feature-selection and dimensionality-reduction principles, explainable decision analysis, and evaluation using classification-oriented performance measures. The framework is conceptually grounded in established studies of feature selection, dimensionality reduction, machine learning, and evaluation metrics. Mutual-information-based feature selection is positioned as a mechanism for reducing redundant industrial telemetry while preserving discriminative information, whereas dimensionality reduction is used to address high-dimensional and heterogeneous observations. Federated aggregation enables collaborative learning while limiting direct data exchange. An explanation layer subsequently identifies the principal features influencing individual security decisions, improving analyst trust and operational interpretability. The resulting framework is particularly relevant to heterogeneous smart factories where data distributions, device capabilities, and attack patterns may vary considerably across participating sites. The analysis demonstrates that explainability, privacy preservation, dimensionality management, and predictive performance should be treated as interconnected design objectives rather than isolated components. The study also identifies limitations associated with heterogeneous client distributions, computational overhead, explanation fidelity, and the absence of direct empirical validation in a deployed industrial environment.

Explainable Artificial Intelligence, Federated Learning, Industrial Internet of Things, Smart Factory

M. Altin and A. Cakir, “Exploring the influence of dimensionality reduction on anomaly detection performance in multivariate time series,” Mar. 2024, arXiv:2403.04429.

M. Beraha, A. M. Metelli, M. Papini, A. Trinzoni, and M. Rostelli, “Feature selection via mutualinformation: New theoretical insight,” in Int. Joint Conf. Neural Netw. (IJCNN), Budapest, Hungary, Jul.

C. M. Bishop, Pattern Recognition and Machine Learning. New York, NY, USA: Springer, 2006.

G. Chandrashekar and F. Sahin, “A survey on feature selection methods,” Comput. Electr. Eng., vol. 40, no. 1, pp. 16–28, Jan. 2014. 10.1016/j.compeleceng.2013.11.024.

CIC-DDoS 2019 Dataset. Accessed: Feb. 5, 2024. [Online]. Available: https://www.unb.ca/cic/datasets/ ddos-2019.html

A. Geron, Hands-on Machine Learning with Scikit-Learn. Oreilly, Boston: Keras & TensorFlow, 2019.

S. A. Hicks et al., “On evaluation metrics for medical applications of artificial intelligence,” Nature Sci.

Rep., vol. 12, no. 5979, pp. 1–9, Apr. 2022. 10.1038/s41598-022-09954-8.

Metric and scoring: quantifying the quality of predictions. Accessed: Mar. 10, 2024. [Online]. Available:

https://scikitlearn.org/stable/modules/generated/sklearn.metrics.precision_recall_fscore_support.html

A. Pal, “Logistic Regression: A Simple Primer,” Cancer Res. Stat. Treat. vol. 4, no. 3, pp. 551–554, Jul.

H. T. Pham, J. Awange, and M. Kuhn, “Evaluation of three feature dimension reduction techniques formachine-learning based crop yield prediction models,” Sensors, vol. 22, no. 17, pp. 1–18, Sep. 2022. 10.3390/s22176609.

J. R. Vergara and P. Estevez, “A review of feature selection methods based on mutual information,” Neural Comput. Appl., vol. 24, no. 1, pp. 10–23, Jan. 2014. 10.1007/s00521-013-1368-0.

K. Ramamurthy, R. K. Konduru and N. Amanmadov, "EvoGraphCoder: An Evolutionary Graph-Reasoning Framework for Self-Adaptive Software Engineering," in IEEE Access, vol. 14, pp. 63063-63076, 2026, doi: 10.1109/ACCESS.2026.3686019.

Philip, P. G. (2024). Artificial Intelligence-Driven Project Risk Prediction Models: Enhancing Decision-Making Accuracy in Large-Scale Infrastructure Projects . American Journal of Technology, 3(1), 52–69. https://doi.org/10.58425/ajt.v3i1.571

Govindarajan, V., Ahmed, F., Kamaluddin, K. et al. SecureRiskNet: An Advanced AI-Driven Framework for Intelligent Security Risk Detection in Heterogeneous Cloud-Fog Computing Networks. Int J Comput Intell Syst 19, 74 (2026).