Communication Dans Un Congrès Année : 2023

Red Team LLM: towards an adaptive and robust automation solution

Dorian Bachelot
Tudy Gourmelen
  • Fonction : Auteur
  • PersonId : 1113000
Adrien Quemat
  • Fonction : Auteur
  • PersonId : 1112999
Pierre-Marie Satre
  • Fonction : Auteur
  • PersonId : 1112991
Loïc Scotto
  • Fonction : Auteur
Di Perrotolo
  • Fonction : Auteur
  • PersonId : 1320943
Maximilien Chaux
  • Fonction : Auteur
  • PersonId : 1320944
Pierre Delesques
  • Fonction : Auteur
  • PersonId : 1320945
Olivier Gesny

Résumé

Artificial intelligence has become really popular in recent years, especially its embedding in cybersecurity applications. Today, studies have shown that reinforcement learning agents are able to find the optimal sequence of actions in order to attack a network. However, these agents are often over-trained and can neither adapt nor be robust to different networks from the ones they were trained on. We propose a new agent based on a zero-shot approach that adapts itself to any given network and that is robust to parameters and objectives changes without requiring another training phase. We introduce a new metric that better measures the ability of agents to attack a network without prior knowledge. In this paper, we also discuss about the first steps towards explainability for our model and its future improvements.

Fichier principal
Vignette du fichier
paper5.pdf (751.5 Ko) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence

Dates et versions

hal-04328468 , version 1 (07-12-2023)

Licence

Identifiants

  • HAL Id : hal-04328468 , version 1

Citer

Christophe Genevey-Metat, Dorian Bachelot, Tudy Gourmelen, Adrien Quemat, Pierre-Marie Satre, et al.. Red Team LLM: towards an adaptive and robust automation solution. Conference on Artificial Intelligence for Defense, DGA Maîtrise de l'Information, Nov 2023, Rennes, France. ⟨hal-04328468⟩

Collections

855 Consultations
703 Téléchargements

Partager

  • More