Ing. Sebastián García, Ph.D.

This dissertation should explore the application of Large Language Models (LLMs) as a fundamental technology in the development of deception strategies within cybersecurity frameworks. The research will encompass a detailed examination of diverse LLMs, including those operating on cloud infrastructures, local models, and fine-tuned models, as well as the training of models tailored for specific security tasks. The primary focus of the thesis will be to assess how LLMs, both autonomously and in conjunction with complementary AI systems, can enhance organizational defense mechanisms. This includes the potential for LLMs to more effectively mimic organizational services and human interactions, strategically manage responses to cyber threats, and facilitate the creation of cost-effective, internal digital twins of entire organizations. The study aims to provide a thorough understanding of the strategic utilization of LLMs to bolster cybersecurity defenses, offering significant insights into their potential and limitations in a high-stakes domain.

The topic is related with the problem of explaining machine learning in network security with adversarial distributed environments. It is common to take security decisions based on information coming from other peers in the network in a lot of environments, but it is hard to explain why we took some decisions and on which data our decision was based on. Moreover, it is hard to know how much of the decision was taken based on adversarial data. This is a common situation in network security decisions.