AI-DRIVEN FEDERATED LEARNING FRAMEWORK FOR PRIVACY-PRESERVING EARLY DETECTION OF CYBER THREATS IN PAKISTAN’S CRITICAL INFRASTRUCTURE SYSTEMS

Abstract

The increasing digitalization of critical infrastructure systems has heightened their vulnerability to sophisticated cyber threats, particularly in developing countries such as Pakistan. This study proposed an AI-driven federated learning (FL) framework for privacy-preserving early detection of cyber threats across distributed infrastructure environments. The framework leveraged decentralized model training to enable collaborative intelligence without sharing sensitive data, thereby addressing privacy and security concerns associated with traditional centralized approaches. A quantitative experimental design was employed, utilizing approximately 120,000 labeled cybersecurity instances distributed across 50 simulated client nodes representing heterogeneous infrastructure systems.

The proposed model integrated deep learning techniques with federated optimization and incorporated privacy-preserving mechanisms, including secure aggregation and differential privacy. Performance evaluation demonstrated that the FL framework outperformed centralized models, achieving higher accuracy (95.6%), precision (94.2%), recall (93.8%), and F1-score (94.0%), while significantly reducing detection latency. Additionally, the framework exhibited strong scalability and enhanced resilience against adversarial attacks when robust aggregation techniques were applied. Privacy analysis confirmed a substantial reduction in data leakage risks due to the elimination of raw data sharing.

The findings highlight the effectiveness of federated learning as a scalable, secure, and privacy-aware approach for cybersecurity in critical infrastructure systems. This study provides a context-specific solution for Pakistan and offers practical and theoretical contributions toward strengthening national cyber resilience through decentralized and intelligent threat detection mechanisms.