Theses
THESIS PRE-DEFENSE -- Contributions to cybersecurity risk management for Artificial Intelligence systems
Speaker: José Manuel Camacho Rodríguez ()Advisor/s: David Ríos Insua (ICMAT-CSIC); Roi Naveiro (CUNEF Universidad)Date: Friday, 23 January 2026 - 14:00Place: Aula Gris 1, ICMAT
Abstract:
Artificial Intelligence (AI) is becoming a core element in modern digital systems, advancing numerous technologies and disciplines, such as drug discovery, autonomous driving, or cybersecurity in significant ways. However, in relation to cybersecurity, AI is also a double-edged sword. Alongside its benefits, AI introduces novel risks that expand the cybersecurity threat landscape. These include emerging vulnerabilities in AI-based systems and diverse impacts associated with AI. As a result, current cybersecurity risk management (CSRM) methodologies, designed for traditional information systems, are limited in addressing the new reality.
This thesis advances the CSRM domain to address the issues emerging from AI-based systems from a holistic perspective through its first objective (O1), while the remaining objectives (O2 to O4) further develop specific components of the field in its intersection with AI. To achieve this, Adversarial Risk Analysis (ARA) is used as a recurring element, as it provides an alternative modeling framework to game theory that enables cybersecurity scenarios to be analyzed more realistically.
The first objective of the thesis (O1), covered in Chapter 2, is to develop a cybersecurity risk analysis framework tailored to systems with AI components. The framework addresses the new challenges posed by AI in risk management and supports the selection of an optimal protection portfolio to mitigate risk. It also employs ARA to model targeted threats affecting the system under analysis. The applicability of the framework is demonstrated through a case study on automated driving systems.
The second objective of the thesis (O2), accomplished in Chapter 3, is to implement a computational methodology that applies ARA to identify optimal decisions in scenarios involving two adversaries (a defender and an attacker) modeled as security games. By contributing to the analysis and solution of these games, O2 supports a deeper understanding of attacks that may target AI systems and potentially improves their protection. The methodology handles multi-stage games and accommodates both discrete and continuous decision domains. Using a disinformation war example, the chapter shows how the methodology can provide optimal defensive decisions and forecast attacker behavior in security contexts.
The third (O3) and fourth (O4) objectives of this thesis examine security controls that rely on AI. The objective O3, addressed in Chapter 4, focuses on investigating the vulnerabilities of Hidden Markov Model (HMM) inference when input batch observation data is corrupted. The chapter introduces a set of problems that target filtering, smoothing, and decoding inferences by adopting an ARA approach to model the malicious actor attempting to manipulate the HMM. To tackle the complexity of these problems, the chapter develops and evaluates three algorithmic routines. Extensive experiments, including a part-of-speech tagging case study, show that even small perturbations to batches of data can substantially degrade HMM inference and prediction.
The fourth objective (O4), achieved in Chapter 5, is to examine machine learning methods for detecting bots in Twitter. The chapter presents a comparative study of multiple supervised detection models, evaluating data requirements and performance across different bot behaviors. Building on these results, the chapter applies one of the models to analyze bot presence during the COVID-19 pandemic and develops an application to support the analysis of content spread by bots.
