CV

Postdoc Fellow

Adelaide University & Australian Institute for Machine Learning

Conducted research across vision language models, LLM safety, and LLM Agents.

• Understand and Improve the reliability, security and safety of AI systems.
• Develop a parameter-efficient few-shot learning algorithm for vision–language foundation models.
• Coordinate Applied Artificial Intelligence and Machine Learning
• Supervise postgraduate research projects in Responsible AI and LLM Safety and in Security Vulnerabilities in Autonomous LLM Agents

PhD

The University of Adelaide

Computer Sciences

• Towards Robust Deep Neural Networks - Query Efficient Black-box Adversarial Attacks and Defences
• Investigated the security of deep learning architectures under Black-Box threat models, simulating real-world deployment constraints.

Master of Engineering

Royal Melbourne Institute of Technology, Ho Chi Minh city ,Vietnam

Computer Engineering

NDSS Student Grant

Network and Distributed System Security (NDSS) Symposium

Awarded for full-time students participating NDSS Symposium.

Postgraduate Research Scholarship

Faculty of Engineering, Computer & Mathematical Sciences

Awarded for talented students worldwide who want to carry out cutting-edge research and make a real impact on the world.

Intel Scholarship

Intel incorporates

Awarded for the brightest and best students who will be the core technical resource for Intel Vietnam and the high-tech industry of Vietnam.

Certified but Fooled! Breaking Certified Defences with Ghost Certificates

AAAI

This study shows how probabilistic certification frameworks for machine learning can be maliciously manipulated. While past work relied on large, noticeable changes, the authors use “region-focused adversarial examples” to craft small, imperceptible perturbations. These effectively spoof safety certificates and bypass state-of-the-art defenses, exposing critical limits in current robustness certification methods.

BruSLeAttack Query-Efficient Score-Based Sparse Adversarial Attack

ICLR

This study addresses the NP-hard, non-differentiable challenge of generating sparse adversarial attacks against black-box models using only score-based query replies. The authors introduce BruSLeAttack, a faster, query-efficient algorithm. It achieves state-of-the-art success rates on ImageNet, effectively bypasses defenses, and exposes vulnerabilities in real-world systems like Google Cloud Vision..

Query efficient decision based sparse attacks against black-box deep learning models

ICLR

This study introduces SparseEvo, an evolution-based algorithm designed for decision-based sparse adversarial attacks against black-box deep learning models, including vision transformers. Addressing an NP-hard optimization challenge, SparseEvo achieves state-of-the-art query efficiency and competitiveness on ImageNet while exposing critical safety vulnerabilities in deployed real-world machine learning systems..