Institution profile

Universiti Teknologi Malaysia

Academic institutionasia · my
Official website
Research library22linked papers
Opportunities0open roles
Selected work

Representative Papers

Privacy-Preserving Dataset Curation for Kuala Lumpur Urban Traffic: Grounded Vision-Language Detection with Spatial Vehicle-Context Filtering

Aug 12, 2026

This study addresses the high privacy risks inherent in complex tropical traffic scenarios in Kuala Lumpur by proposing an automated anonymization framework. By integrating Grounding DINO with a novel spatial vehicle ROI constraint mechanism, temporal persistence, and automated quality inspection, the method effectively suppresses environmental false positives and enhances occluded object recognition while preserving scene context. Experimental evaluation on a 1,266-frame test set demonstrates an anonymization success rate of approximately 95%, significantly mitigating privacy leakage risks. Consequently, this approach provides a robust solution for the efficient de-identification of urban traffic datasets in tropical environments, balancing data utility with stringent privacy protection requirements.

0 citationsRead paper

Certifying Collective Reasoning in Multi-Agent Systems via Koopman Spectral Analysis

Aug 06, 2026

This work addresses the lack of verifiable guarantees—such as convergence, interpretability, and bounded interaction rounds—in multi-agent large language model reasoning. The authors model collective reasoning as a nonlinear dynamical system over a communication graph and, for the first time, apply Koopman operator theory to construct a linear representation from interaction trajectories. Spectral analysis of this representation yields three machine-verifiable certificates: convergence deadlines, identification of cohesive cliques with interpretable validity, and an auditable basis for compressed messages. Experiments demonstrate that convergence rounds are predicted accurately in 96% of configurations (log-scale correlation of 0.93), attributions are exact, and decision-relevant information is preserved with 99.7% fidelity using only 8 out of 32 spectral coordinates. Certificates trained on 15 debates remain fully valid across 60 leave-one-out tests and are computable within minutes on a CPU.

0 citationsRead paper

LLM-Guided Initialization for Accelerated Hybrid Quantum-Classical Medical Image Classification

Jul 29, 2026

This work addresses the challenge of barren plateaus in variational quantum algorithms for medical image classification, which lead to vanishing gradients and hinder training. The authors propose a novel approach that integrates large language model (LLM)-guided single-query initialization with AdaInit, CUDA-Q GPU-accelerated quantum simulation, and prompt engineering to generate high-quality initial parameters without iterative optimization. Applied to binary classification of mammogram images, this method avoids barren plateaus effectively, yielding a 14.6× increase in gradient variance and a 160× acceleration in convergence time (1.1 seconds versus 176 seconds) compared to random initialization, while maintaining a classification accuracy of 61.4%. These results demonstrate a significant improvement in the trainability and efficiency of hybrid quantum-classical models.

0 citationsRead paper
Recent publications

Latest Papers

Privacy-Preserving Dataset Curation for Kuala Lumpur Urban Traffic: Grounded Vision-Language Detection with Spatial Vehicle-Context Filtering

Aug 12, 2026

This study addresses the high privacy risks inherent in complex tropical traffic scenarios in Kuala Lumpur by proposing an automated anonymization framework. By integrating Grounding DINO with a novel spatial vehicle ROI constraint mechanism, temporal persistence, and automated quality inspection, the method effectively suppresses environmental false positives and enhances occluded object recognition while preserving scene context. Experimental evaluation on a 1,266-frame test set demonstrates an anonymization success rate of approximately 95%, significantly mitigating privacy leakage risks. Consequently, this approach provides a robust solution for the efficient de-identification of urban traffic datasets in tropical environments, balancing data utility with stringent privacy protection requirements.

0 citationsRead paper

Certifying Collective Reasoning in Multi-Agent Systems via Koopman Spectral Analysis

Aug 06, 2026

This work addresses the lack of verifiable guarantees—such as convergence, interpretability, and bounded interaction rounds—in multi-agent large language model reasoning. The authors model collective reasoning as a nonlinear dynamical system over a communication graph and, for the first time, apply Koopman operator theory to construct a linear representation from interaction trajectories. Spectral analysis of this representation yields three machine-verifiable certificates: convergence deadlines, identification of cohesive cliques with interpretable validity, and an auditable basis for compressed messages. Experiments demonstrate that convergence rounds are predicted accurately in 96% of configurations (log-scale correlation of 0.93), attributions are exact, and decision-relevant information is preserved with 99.7% fidelity using only 8 out of 32 spectral coordinates. Certificates trained on 15 debates remain fully valid across 60 leave-one-out tests and are computable within minutes on a CPU.

0 citationsRead paper

LLM-Guided Initialization for Accelerated Hybrid Quantum-Classical Medical Image Classification

Jul 29, 2026

This work addresses the challenge of barren plateaus in variational quantum algorithms for medical image classification, which lead to vanishing gradients and hinder training. The authors propose a novel approach that integrates large language model (LLM)-guided single-query initialization with AdaInit, CUDA-Q GPU-accelerated quantum simulation, and prompt engineering to generate high-quality initial parameters without iterative optimization. Applied to binary classification of mammogram images, this method avoids barren plateaus effectively, yielding a 14.6× increase in gradient variance and a 160× acceleration in convergence time (1.1 seconds versus 176 seconds) compared to random initialization, while maintaining a classification accuracy of 61.4%. These results demonstrate a significant improvement in the trainability and efficiency of hybrid quantum-classical models.

0 citationsRead paper