Senior Lecturer (US Associate Professor equivalent) · British University Vietnam
I build machine learning systems that stay reliable when they are attacked, audited or deployed: intrusion and malware detection, graph learning, privacy-preserving training and verifiable machine unlearning.
PhD in Computer Science (Université de Lorraine, Inria Nancy). Before returning to academia I led data science teams in fintech, e-commerce and digital banking, so I care about methods that survive contact with real data.
My work has moved from trust and collaboration in online systems towards secure, explainable and privacy-aware AI. Papers can belong to more than one theme.
Machine learning to detect intrusions, malware and spam in IoT, SDN and WiFi environments.
Publications →Graph neural networks, community detection and link prediction for fraud, bots and malware.
Publications →Federated, Byzantine-robust and privacy-preserving learning, and certifying machine unlearning.
Publications →Explainability, uncertainty, fairness and conformal methods for models people can trust in operation.
Publications →Retrieval-augmented generation, transformers and multi-agent LLMs for search, finance and security.
Publications →Computational trust, Wikipedia quality assessment and social recommendation (doctoral work at Inria).
Publications →Fraud detection, financial documents, trading and portfolio decision-making.
Publications →Work-integrated learning, assistive AI and equity in higher education and school outreach.
Publications →Plants paired ghost canaries before training and uses betting statistics that stay valid however long an auditor keeps watching, so a deletion claim can be certified rather than trusted.
F1 of 0.9474 on TwiBot-20, 9.64% above recent baselines, with a forensic explanation module that moderators can act on.
Federated graph learning for fraud detection that treats privacy and group fairness as design constraints.
Combines nine SPROUT uncertainty measures with distance-based metrics and Isolation Forest to cut false negative rates on minority attack classes by up to 50% on CICIDS 2017, with accuracy above 98%.
Retrieval-augmented generation over long, table-heavy financial documents.
Distribution-free, conformal inference for causal effects when competing risks are present.
From my doctoral research on computational trust in large-scale collaborative systems.
Publications per year, from the ORCID record.
I am a data science leader and senior lecturer. My research record spans AI, cyber security and collaborative systems, and my teaching covers data science, machine learning and NLP at British University Vietnam.
I am most interested in models that do more than score well on a benchmark. They need to be explainable, robust to adversaries and verifiable by someone who does not trust the people who trained them.
Open to collaboration on trustworthy ML, graph learning, federated learning and AI in education. The quickest way to reach me is through LinkedIn or the BUV faculty page.