Secure Data Science Model with Cryptographic Techniques for Cybersecurity Risk Mitigation via Machine Learning

Abstract

The rapid expansion of the Industrial Internet of Things (IIoT) has introduced complex cyber security challenges,
necessitating intelligent and privacy-preserving intrusion detection solutions. This paper presents an Intelligent Digital Twin-based Federated Learning (IDT-FL) framework that employs probabilistic feature computation and distributed model training for secure and efficient cyber attack detection. The framework incorporates information gain-based probabilistic feature ranking, adaptive local model updates, and global aggregation to enhance detection accuracy while maintaining data locality. Extensive experiments on benchmark datasets—CICIDS2017, N-BaIoT, and ToN_IoT—demonstrate that IDT-FL consistently outperforms conventional Federated Averaging and centralized DNN models. Specifically, IDT-FL achieves an average accuracy of 95.0%, F1-score of 94.9%, and latency as low as 20.1 ms across datasets, with reduced communication overhead (average 32.9 MB) and faster convergence (average 40 rounds). In comparison, traditional federated models report lower accuracy (avg. 91.9%) and higher latency (avg. 25.5 ms). The proposed IDT-FL framework not only ensures high detection precision but also optimizes network efficiency and scalability, making it a viable solution for real-time and distributed cyberattack detection in IIoT-enabled Cyber-Physical Systems.