We are honored to invite you to Julien Soule’s PhD defense, entitled: “On the Organisation of a Multi-Agent Cyberdefense System”.
The defense will take place on Monday, November 17th, 2025 at 10:00 a.m at the LCIS laboratory (Univ. Grenoble Alpes / Grenoble INP).
It will be held in French and is open to the public.
📄 The manuscript and further details are available online on this webpage.
In the face of increasingly complex cybersecurity threats, centralized approaches show their limits in protecting distributed and dynamic systems. This thesis explores a distributed approach based on Multi-Agent Systems (MAS) capable of collectively detecting, responding, and adapting to autonomous and evolving attacks.
The central goal is to enable the design of a cyber defense MAS by identifying an organizational mechanism adapted to both designers’ constraints and environmental dynamics. The literature highlights a symbolic approach fostering control, and a connectionist approach favoring performance. To overcome this tension, the thesis proposes a hybrid method combining a symbolic organizational model with multi-agent reinforcement learning (MARL).
The key idea is to frame the design of a MAS as a constrained optimization problem, in which the joint policy of agents is learned while respecting organizational constraints expressing the designer’s requirements. This approach requires both faithful modeling of the target environment and the ability to analyze and control the resulting behaviors.
The proposed methodology integrates four activities: (i) modeling the target environment using either manual techniques or World Models, to obtain a simulated version of the target system; (ii) training the agents via MARL, with constraints derived from the organizational model MOISE+MARL+; (iii) analyzing the learned policies, by extracting implicit roles and goals through unsupervised methods applied to trajectories; (iv) transferring the results into the real environment, with continuous updates of both models and policies.
A software platform was developed to implement this methodology and was applied to three use cases: a drone swarm, a company infrastructure, and a microservices architecture. The results show improvements in resilience, adaptability, and autonomy compared to centralized approaches.
Finally, the thesis opens several research perspectives: enhancing environment modeling through expert knowledge integration, improving the robustness of learning in dynamic environments, and exploring latent representations to support organizational analysis.
The thesis was prepared within the framework of the Cyb’Air chair jointly with La Ruche (Thales LAS), which provided an application framework and technological support in AI, as well as within the LCIS laboratory (CO⁴SYS team), which provided expertise in MAS and collective AI.
📌 Zoom link: https://grenoble-inp.zoom.us/j/91692482911
📌 Zoom password: 747546
The defense is open to the public and will take place at the LCIS laboratory, 50 rue Barthelemy de Laffemas, 26000 Valence, in amphitheater D030 (to the right at the building entrance).
No posts found!
© Copyright 2012 – 2023 | Mentions Légales | Tous droits réservés
