Tung Vu

Tung Vu

Auto-generated profile, based on publicly available information

Tung Vu

Posts and Telecommunications Institute of Technology (PTIT), Hanoi, Vietnam

Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
13 May 2025
Verified
Book Chapter
Closed access
PNPhuong Anh Nguyen
Phuong Anh Nguyen
TVTung Son Vu
Tung Son Vu
TBThuan Minh Bui
Thuan Minh Bui
ALAnh Ngoc Le
Anh Ngoc Le
In the rapidly evolving field of video surveillance and security, there is an urgent need for facial recognition systems that can autonomously adapt and improve their accuracy in real-time, particularly in dynamic environments where new faces frequently appear, and lighting conditions constantly change. This paper addresses the critical challenge of developing self-learning camera systems capable of continuously enhancing facial recognition performance without manual intervention, a requirement that has become increasingly vital for maintaining effective security measures in various sectors including public safety, retail, and smart cities. To tackle this pressing issue, we propose a novel autolearning method that combines RetinaFace for face detection and ArcFace for face recognition, incorporating a self-updating mechanism. Our approach consists of three key steps: (1) automatic collection of high-confidence facial images (>90% accuracy) during recognition tasks, (2) periodic model retraining using the collected data, and (3) iterative model updating to enhance overall accuracy.

10 April 2026 in International Journal of Online and Biomedical Engineering (iJOE)
Verified
Journal Article
Open access
PDF file
TVTung Vu
Tung Vu
KTKien Luong Trung
Kien Luong Trung
TBThuan Bui
Thuan Bui
MDMinh Quang Dang
Minh Quang Dang
PNPhuong Anh Nguyen
Phuong Anh Nguyen
NLNgoc Le
Ngoc Le
Inefficient non-clinical information delivery in medical centers burdens patients and staff, despite the potential of LLM-enhanced humanoid robots. Practical deployment faces critical hurdles: high cloud-LLM latencies, on-robot computational limits, and lacking secure, distributed knowledge sharing under strict privacy regulations. This paper introduces Dynamic LLM Routing for Smart Medical Robots, a framework designed for clinical question-answering systems. Our hybrid edge-cloud architecture integrates three core innovations: a dynamic LLM selection mechanism routing queries based on context (complexity, domain, and urgency), an optimized edge computing layer for privacy-preserving local processing, and an intelligent, privacy-preserving caching system enabling secure knowledge sharing and continuous learning. Evaluated on 7,500 authentic hospital queries, our system achieved a 50% reduction in average response latency (0.62s vs. 1.23s), an 89.7% completion rate, and a 62% reduction in external data transmission (73% cache hit rate). These results demonstrate our framework’s efficacy in enabling scalable, efficient, and privacy-compliant AI-powered humanoid robots for critical healthcare information delivery.

23 June 2026 in Computer Vision and Image Understanding
Verified
Journal Article
Closed access
TVTung Vu
Tung Vu
HNHai Nguyen
Hai Nguyen
CTCong Tran
Cong Tran

26 April 2025
Verified
Book Chapter
Closed access
PDF file
TVTung Vu
Tung Vu
HNHung Nguyen
Hung Nguyen
NNNam Nguyen
Nam Nguyen
CPCuong Pham
Cuong Pham
CTCong Tran
Cong Tran

29 May 2026 in IEEE Access
Verified
Journal Article
Open access
TVTung Vu
Tung Vu
LNLam Nguyen
Lam Nguyen
QDQuynh Dao
Quynh Dao
Large Language Models (LLMs) have demonstrated remarkable capabilities in generating synthetic text data, offering transformative potential for data augmentation, privacy preservation, and addressing data scarcity across sensitive domains. However, when the subjects of generated content involve populationsincluding children, patients, refugees, racial and ethnic minorities, LGBTQ+ individuals, and persons with disabilitiescritical ethical risks emerge, spanning privacy breaches, bias amplification, stereo-typical representation, and the production of harmful content. Despite growing interest in both synthetic data generation and AI safety, there remains no comprehensive survey that systematically examines the intersection of ethical synthetic text generation and vulnerable population protection. This paper addresses this gap by providing the first comprehensive survey of methods, frameworks, evaluation metrics, and open challenges in this domain. We review over 300 works spanning synthetic data generation, prompt engineering, LLM safety and alignment, fairness and bias mitigation, and privacypreserving techniques. We propose a taxonomy that classifies existing approaches along three axes: the generation pipeline stage (pregeneration, ingeneration, postgeneration), the ethical dimension addressed (privacy, fairness, safety, representation), and the target population. We critically analyze four systematic failure patterns in existing methodspopulation blindness, intersectionality gaps, evaluation circularity, and static ethical requirementsand provide a research gaps matrix mapping understudied populationdimension combinations. Furthermore, we introduce PromptGuard, a modular orchestrated prompting framework that integrates existing prompting techniquesincluding ChainofThought, TreeofThoughts, Constitutional AI, ReAct, and SelfCritiqueinto a principled sixmodule architecture designed for populationspecific ethical text generation. We demonstrate PromptGuard through three detailed case studies spanning LGBTQ+ youth mental health, clinical records, and refugee narratives. We additionally analyze the regulatory landscape (EU AI Act, COPPA, HIPAA), crosscultural ethical tensions, and model collapse risks. Finally, we distill practical guidelines for practitioners and outline future research directions toward more inclusive, culturally adaptive, and communitydriven ethical AI systems.

22 July 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
VTVu Tung
Vu Tung
DLDung Le
Dung Le
HPHuy Phan
Huy Phan
HNHieu Nguyen
Hieu Nguyen
ANAn Nguyen
An Nguyen
QDQuynh Dao
Quynh Dao
Vision foundation models based on the Vision Transformer (ViT) architecture have emerged aspowerful feature extractors deployed across diverse downstream tasks, yet they remain acutely vulnerable to adversarial perturbations that destabilize their representations even under imperceptible input modifications. Existing defenses-most notably adversarial training and certified randomized smoothing-achieve robustness at the cost of retraining or substantially modifying the backbone, impractical for large multi-task foundation models. We propose Robustness Tokens, a family of 𝜅 learnable token vectors appended to a frozen ViT backbone’s input sequence, optimized via a lightweight dual objective that preserves clean features while suppressing adversarial feature drift. Training takes under 300 gradient steps on a small ImageNet subset and updates a negligible fraction of parameters, leaving the backbone untouched. Across eleven architectures from two ViT families-DINOv2 and DeiT-III-Robustness Tokens raise feature-space cosine similarity under a non-adaptive PGD threat model above 0.82, versus undefended baselines as low as 0.011, with limited clean-performance degradation. Under an adaptive white-box threat model, cosine similarityfor DINOv2-B drops to 0.309, still well above the undefended baseline (0.044) but far from a full guarantee. Extending the adaptive evaluation to downstream tasks shows this gap is task-dependent: classification accuracy under adaptive attack collapses to 0.7%, indistinguishable from baseline, while segmentation retains a partial, order-of-magnitude benefit (27.1 vs. 5.2 mIoU). These results establish a parameter-efficient mechanism for hardening frozen foundation models against non-adaptive adversaries across two ViT pre-training paradigms, while leaving full white-box robustness,especially for classification, an open problem.