Sécurité des réseaux 5G/6G par l’IA : Prédiction de la durée des attaques pour une mitigation intelligente

Abstract

The rapid proliferation of 5G and O-RAN network architectures has dramatically expanded the cyber threat landscape, creating critical security challenges that demand innovative solutions beyond conventional intrusion detection systems. This research introduces an innovative framework for cyberattack duration prediction that transforms cybersecurity from reactive detection to proactive threat management, addressing the fundamental gap in temporal threat intelligence that currently limits effective mitigation strategies in next-generation networks. The framework is built upon three key contributions : an advanced Transformer-based neural architecture specifically engineered for capturing complex temporal patterns in attack sequences, demonstrating significant performance improvements over traditional recurrent models ; a privacy-preserving federated learning paradigm with adaptive client selection and dynamic training optimization that enables intelligent resource allocation across distributed network nodes while maintaining data confidentiality and computational efficiency ; and the novel FML-AD hybrid framework that synergistically combines federated learning with meta learning techniques to overcome domain adaptation barriers between training environments and real-world operational deployment. The adaptive client selection mechanism strategically prioritizes underperforming nodes based on real-time validation metrics, while the dynamic training optimization adjusts epoch allocation and batch processing according to client capabilities and data distributions. Extensive experimental validation conducted on multiple benchmark datasets and a sophisticated 5G O-RAN testbed demonstrates exceptional prediction accuracy for remaining attack duration, facilitating optimized resource allocation, intelligent mitigation planning, and guaranteed quality of service for legitimate users while maintaining robust privacy safeguards across heterogeneous network infrastructures. This research establishes the foundational principles for next-generation adaptive cybersecurity systems capable of proactive threat anticipation and dynamic defense optimization, with promising future research directions encompassing multi-attack scenario handling, energy aware model optimization, and real-time inference engines for large-scale industrial deployments, ultimately contributing to enhanced resilience of critical digital infrastructures against evolving cyber threats in the 5G era and beyond.

Date	15 May 2026
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Rami Langar (Supervisor) & Waël Jaafar (Co-supervisor)