Apprentissage par renforcement pour une cybersécurité résiliente

Abstract

As global reliance on IoT devices and interconnected technologies intensifies, cybersecurity threats are evolving at an unprecedented pace. Traditional rule-based defense mechanisms and even conventional machine learning models have shown diminishing effectiveness against the dynamic and adaptive nature of modern cyber threats. To address this gap, this thesis introduces a novel RL based framework aimed at enhancing cybersecurity resiliency. Specifically, we propose the use of RL agents to simulate adversarial behavior by launching adaptive cyberattacks and evaluating the robustness of target networks under varying conditions. The study begins by positioning reinforcement learning within the broader landscape of artificial intelligence in cybersecurity, rule-based approaches. We then design and implement a reinforcement learning model integrated with model-agnostic meta-learning to enable rapid adaptation and generalization across heterogeneous network environments. Unlike previous works that are limited to narrowly scoped settings, our RL model demonstrates generalization capability, enabling more effective penetration testing. Our experimental results show that reinforcement learning, augmented with meta-learning, provides a scalable and adaptive methodology for offensive and cybersecurity operations. By empowering systems to learn from cyber interactions, this approach contributes a resilient, intelligent layer to modern cybersecurity architectures.

Date	17 Dec 2025
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Chamseddine Talhi (Supervisor) & Hakima Ould-Slimane (Co-supervisor)