Ciberseguridad en la era de la IA generativa y la IA explicable
Loading...
Official URL
Full text at PDC
Publication date
2025
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citation
Abstract
Este trabajo tiene como objetivo crear un chat de inteligencia artificial entrenado con datos de ciberseguridad, para así apoyar al aprendizaje de toda aquella persona que esté interesada en realizar máquinas de VulnHub y aprender
sobre el ámbito de la seguridad informática. Para ello decidimos entrenar un modelo de inteligencia artificial con el modelo DeepSeek-7B como base, y apoyarlo con un dataset realizado por nosotros sobre la resolución de máquinas en VulnHub. El desarrollo de este dataset se puede dividir en dos fases:
• Primera fase: Elaboración y resolución de máquinas de VulnHub, para desarrollar los walkthrough.
• Segunda fase: Conversión de estos documentos a un dataset valido para el entrenamiento del modelo.
Tras esta conversión para crear un dataset adecuado, se entrenó al modelo varias veces hasta conseguir uno que cumpliera los requisitos del equipo. Para poder realizar las pruebas del modelo este se subió al portal web local OpenWebUI, ya que su interfaz es bastante intuitiva.
Por último, decidimos crear dos descargables con Docker:
• Modelo 7B: más potente, pero más complicado de ejecutar en un entorno local “común”.
• Modelo 1.5B: más liviano para poder ser ejecutado en un entorno local “común”.
Con estos dos descargables conseguimos que varias personas nos dieran retrospectiva sobre nuestro chat, obviamente teniendo en cuenta los diferentes modelos, ya que el que nosotros hemos perfeccionado es el 7B. Pero en general los comentarios fueron buenos.
This project aims to create an artificial intelligence chatbot trained with cybersecurity data to support the learning process of anyone interested in solving VulnHub machines and gaining knowledge in the field of information security. To achieve this, we decided to train an AI model using DeepSeek-7B as the base model, supported by a custom dataset we created focused on the resolution of VulnHub machines. The development of this dataset can be divided into two phases: • First phase: Solving and documenting VulnHub machines to create walkthroughs. • Second phase: Converting these documents into a valid dataset for model training. After converting the data into a suitable format, the model was trained several times until we obtained a version that met the team’s requirements. To test the model, we uploaded it to the local web portal OpenWebUI, as its interface is quite intuitive. Finally, we decided to create two Docker-ready downloads: • 7B model: More powerful, but more complex to run in a typical local environment. • 1.5B model: Lighter, making it easier to run on a typical local setup. With these two downloads, we were able to gather feedback from several users about our chatbot, taking into account the differences between the models, since the 7B version is the one we refined. Overall, the feedback was positive.
This project aims to create an artificial intelligence chatbot trained with cybersecurity data to support the learning process of anyone interested in solving VulnHub machines and gaining knowledge in the field of information security. To achieve this, we decided to train an AI model using DeepSeek-7B as the base model, supported by a custom dataset we created focused on the resolution of VulnHub machines. The development of this dataset can be divided into two phases: • First phase: Solving and documenting VulnHub machines to create walkthroughs. • Second phase: Converting these documents into a valid dataset for model training. After converting the data into a suitable format, the model was trained several times until we obtained a version that met the team’s requirements. To test the model, we uploaded it to the local web portal OpenWebUI, as its interface is quite intuitive. Finally, we decided to create two Docker-ready downloads: • 7B model: More powerful, but more complex to run in a typical local environment. • 1.5B model: Lighter, making it easier to run on a typical local setup. With these two downloads, we were able to gather feedback from several users about our chatbot, taking into account the differences between the models, since the 7B version is the one we refined. Overall, the feedback was positive.
Description
Trabajo de Fin de Grado en Ingeniería del Software, Facultad de Informática UCM, Departamento de Ingeniería de Software e Inteligencia Artificial, Curso 2024/2025.
Todo nuestro código realizado, tanto para el entrenamiento de la IA, como para la creación de dataset, y todos los walkthroughs creados, están recopilados en el siguiente GitHub:
https://github.com/Ciberseguridad-con-IA/Trabajo-de-fin-de-grado













