Please note: This PhD defence will take place in DC 1331 and online.
Seyed Soheil Johari, PhD candidate
David R. Cheriton School of Computer Science
Supervisor: Professor Raouf Boutaba
The increasing adoption of Network Functions Virtualization (NFV), cloud-native architectures, and microservice-based designs has transformed modern communication networks into highly dynamic and complex software-driven systems. While these advancements enable flexibility, scalability, and efficient resource utilization, they also introduce significant challenges for network fault and performance management due to increased system complexity, dynamic behavior, and limited observability. In such environments, faults often originate in different layers of the system and propagate across tightly coupled components, manifesting as correlated anomalies in high-dimensional telemetry data. Traditional rule-based and threshold-driven management approaches are no longer sufficient to capture these complex dependencies or reliably diagnose faults, motivating the adoption of data-driven and machine learning (ML)-based techniques for automated network management.
This thesis develops data-driven methods for fault detection, diagnosis, and root cause analysis in virtualized network systems, with a focus on NFV-based infrastructures and cloud-native mobile networks. The central goal is to design learning-based solutions that remain robust under limited supervision, data contamination, evolving environments, and incomplete observability, while providing interpretable and actionable insights for real-world network operations. To this end, the proposed approaches combine unsupervised learning, active learning, domain adaptation, and causal reasoning to address fundamental challenges in modern network management.
Specifically, the thesis includes four main contributions. First, an unsupervised anomaly detection and localization framework for NFV systems that remains robust under contaminated training data by leveraging a teacher-student learning paradigm. Second, an active learning approach for Transformer-based fault diagnosis that reduces labeling costs by selecting informative and diverse samples based on attention-derived dependency structures. Third, a few-shot domain adaptation framework that mitigates data drift through causal feature separation, enabling robust cross-domain deployment without the need for frequent retraining. Fourth, a causal root cause analysis framework that incorporates structural priors learned from normal-operation telemetry and enables the identification of latent root causes under partial observability.
Extensive evaluations on realistic datasets collected from NFV testbeds, cloud-native microservice applications, and 5G network environments demonstrate that the proposed methods significantly improve anomaly detection and localization accuracy, reduce labeling requirements, enhance cross-domain robustness, and provide more accurate and interpretable root cause analysis compared to existing approaches.
To attend this PhD defence in person, please go to DC 1331. You can also attend virtually on MS Teams.