ARTIFICIAL INTELLIGENCE FOR CYBERSECURITY IN DISTRIBUTED SYSTEMS

The rise of distributed systems, encompassing cloud computing, mobile devices, and Internet of Things, has revolutionized modern digital infrastructure but also exposed it to numerous cybersecurity threats. Artificial intelligence (AI), particularly machine learning, has emerged as a powerful tool to enhance security in these distributed environments by detecting and responding to attacks.However, traditional cybersecurity measures are no longer sufficient to address modern threats that constantly evolve to avoid detection. Thus, there is a growing need for adaptive, intelligent security solutions that can keep pace with the dynamic nature of cyber threats.This dissertation presents a comprehensive study of AI techniques to secure networks and networked devices, focusing on the challenge of ensuring that security systems remain effective in the face of evolving threats.The main hurdle in this task is obtaining adequate up-to-date training data in a timely fashion to build robust models.To this end, unsupervised learning methods are explored to detect network intrusions, with a focus on collaborative approaches that leverage the capabilities of multiple devices.For supervised malware classification tasks, Federated Learning (FL) is identified as a promising approach to enable crowdsourced security solutions while preserving the participants' privacy.In addition, novel approaches are developed to ensure that the model performance remain robust over time, even when the data distribution changes.Þspite the potential of FL, the lack of oversight on client behavior can lead clients to deviate from the prescribed learning protocol to obtain unfair advantages.Thus, this dissertation also analyzes the impact of these clients on the model's performance and fairness, and proposes an approach to mitigate their influence,and an incentive mechanism to align the client goals with the server is presented.Additionally, a personalization mechanism is introduced to ensure that each participant obtains a model well-suited to their current local data distribution at any given time.