Unsupervised intrusion detection with cross-domain artificial intelligence methods

• Autores: Rafael San Miguel Carrasco
• Directores de la Tesis: Miguel Ángel Sicilia Urbán (dir. tes.)
• Lectura: En la Universidad de Alcalá ( España ) en 2021
• Idioma: español
• Tribunal Calificador de la Tesis: Rafael Palacios Hielscher (presid.), María Elena García Barriocanal (secret.), Javier Bermejo Higuera (voc.)
• Programa de doctorado: Programa de Doctorado en Ingeniería de la Información y del Conocimiento por la Universidad de Alcalá
• Materias:
  • Matemáticas
    • Ciencia de los ordenadores
      • Inteligencia artificial
• Enlaces
  • Tesis en acceso abierto en: TESEO
• Resumen

  • Cybercrime is a major concern for corporations, business owners, governments and citizens, and it continues to grow in spite of increasing investments in security and fraud prevention. The main challenges in this research field are: being able to detect unknown attacks, and reducing the false positive ratio. The aim of this research work was to target both problems by leveraging four artificial intelligence techniques.

    The first technique is a novel unsupervised learning method based on skip-gram modeling. It was designed, developed and tested against a public dataset with popular intrusion patterns. A high accuracy and a low false positive rate were achieved without prior knowledge of attack patterns.

    The second technique is a novel unsupervised learning method based on topic modeling. It was applied to three related domains (network attacks, payments fraud, IoT malware traffic). A high accuracy was achieved in the three scenarios, even though the malicious activity significantly differs from one domain to the other.

    The third technique is a novel unsupervised learning method based on deep autoencoders, with feature selection performed by a supervised method, random forest. Obtained results showed that this technique can outperform other similar techniques.

    The fourth technique is based on an MLP neural network, and is applied to alert reduction in fraud prevention. This method automates manual reviews previously done by human experts, without significantly impacting accuracy.