Algorithmes d'apprentissage automatique en ligne pour une cyberdéfense optimale dans les grands réseaux // Online machine learning algorithms for optimal cyber-defense in large networks

Ce projet de thèse vise à répondre à cette problématique en proposant des algorithmes efficaces, s'appuyant sur des techniques récentes d'apprentissage automatique (comme les marches aléatoires contrôlées combinées à l'apprentissage par renforcement). L'objectif est de développer des méthodes capables de détecter automatiquement les points faibles d'un réseau, même à grande échelle, afin de renforcer sa résilience face aux cybermenaces.
------------------------------------------------------------------------------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------------------------------------------------------------------------------

Cyberattacks are becoming more and more frequent and increasingly sophis-
ticated, making the automation of defense strategies a priority. One strategy
to address the challenges behind automation in such contexts is by consider-
ing an optimisation-centric approach. Estimating spectrum properties of large
scale graphs is a very hot topic in Computer, Network Science communities
and Machine Learning. In particular, the connectivity of large scale networks is
essential for the study of their performance and resilience against cyberattacks.
Especially, in the context of overlay networks and ad-hoc wireless networks
such understanding is critical.
Machine learning algorithms based on the Power iteration or the Rayleigh quo-
tient techniques have been used to estimate graph spectral properties. When
the graph is unknown, these techniques have been considered recently in
by coupling with a random walk exploration of the graph. In a cybersecurity
context, to know which node and link is a weak point of the network is very
important. Indeed, such vulnerability can lead to successfull cyberattacks on a
network. And if the network is large, it is very complex to identify such weak
point of the structure. This PhD project aims to answer this question and to
propose efficient algorithms based on recent Machine Learning techniques (con-
trolled random walks combined with reinforcement learning for example).
------------------------------------------------------------------------------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------------------------------------------------------------------------------

Début de la thèse : 01/10/2026

Contrat doctoral

Avignon Université

536 Agrosciences et Sciences

Les candidatures doivent avoir des compétences en mathématiques appliquées (modélisation stochastique, optimisation, graphes) et en informatique (python ou matlab) pour du calcul scientifique et simulations.
Ideal candidates will have a strong background in applied mathematics (particularly in stochastic modeling, optimization, and graph theory) as well as experience in scientific programming (Python or MATLAB) for algorithm implementation and simulations.